Bump the version (#2627)

* Allow None input to compiled functions

* Allow None input to compiled functions

.circleci/config.yml

@@ -18,13 +18,14 @@ jobs:
         type: boolean
         default: false
     macos:
-      xcode: "16.2.0"
+      xcode: "26.0.0"
-    resource_class: m2pro.medium
+    resource_class: m4pro.medium
     steps:
       - checkout
       - run:
           name: Install
           command: |
+            xcodebuild -downloadComponent MetalToolchain
             brew install python@3.9
             brew install doxygen
             python3.9 -m venv env
@@ -89,6 +90,7 @@ jobs:
           command: |
             uv venv
             uv pip install cmake
+            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
               uv pip install -e ".[dev]" -v
       - run:
           name: Generate package stubs
@@ -118,7 +120,7 @@ jobs:
     parameters:
       xcode_version:
         type: string
-        default: "16.2.0"
+        default: "26.0.0"
       macosx_deployment_target:
         type: string
         default: ""
@@ -126,12 +128,13 @@ jobs:
       xcode: << parameters.xcode_version >>
     environment:
       MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    resource_class: m2pro.medium
+    resource_class: m4pro.medium
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
+            xcodebuild -downloadComponent MetalToolchain
             HOMEBREW_NO_AUTO_UPDATE=1 HOMEBREW_NO_INSTALL_CLEANUP=1 \
               brew install openmpi uv
       - run:
@@ -196,7 +199,7 @@ jobs:
           name: Run Python tests with JIT
           command: |
             CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              uv pip install -e .
+              uv pip install -e . -v
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
               METAL_DEBUG_ERROR_MODE=0 \
               uv run --no-project python -m xmlrunner discover \
@@ -222,15 +225,20 @@ jobs:
             sudo apt-get update
             sudo apt-get install libcudnn9-dev-cuda-12
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get install libnccl2 libnccl-dev
             curl -sL https://github.com/ccache/ccache/releases/download/v4.11.3/ccache-4.11.3-linux-x86_64.tar.xz | tar xJf -
             sudo mv ccache-4.11.3-linux-x86_64/ccache /usr/bin/ccache
             rm -rf ccache-4.11.3-linux-x86_64
             curl -LsSf https://astral.sh/uv/install.sh | sh
+      - run:
+          name: Set CCache size
+          command: ccache --max-size 1G
       - run:
           name: Install Python package
           command: |
             uv venv
-            CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+            uv pip install cmake
+            DEBUG=1 CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_COMPILE_WARNING_AS_ERROR=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
               uv pip install -e ".[dev]" -v
       - run:
           name: Run Python tests
@@ -238,12 +246,23 @@ jobs:
             source .venv/bin/activate
             LOW_MEMORY=1 DEVICE=cpu python -m unittest discover python/tests -v
             LOW_MEMORY=1 DEVICE=gpu python -m tests discover python/tests -v
+      - run:
+          name: Build CPP only
+          command: |
+            source .venv/bin/activate
+            cmake . -B build \
+              -DMLX_BUILD_CUDA=ON \
+              -DCMAKE_CUDA_COMPILER=`which nvcc` \
+              -DCMAKE_BUILD_TYPE=DEBUG
+            cmake --build build -j `nproc`
+      - run:
+          name: Run CPP tests
+          command: ./build/tests/tests -sfe="*fft_tests.cpp,*linalg_tests.cpp"
       - run:
           name: CCache report
           command: |
             ccache --show-stats
             ccache --zero-stats
-            ccache --max-size 400MB
             ccache --cleanup
       - save_cache:
           key: cuda-<< parameters.image_date >>-{{ arch }}-{{ epoch }}
@@ -257,7 +276,7 @@ jobs:
         default: "3.9"
       xcode_version:
         type: string
-        default: "16.2.0"
+        default: "26.0.0"
       build_env:
         type: string
         default: ""
@@ -266,7 +285,7 @@ jobs:
         default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: m2pro.medium
+    resource_class: m4pro.medium
     environment:
       MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
     steps:
@@ -274,11 +293,15 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            brew install python@<< parameters.python_version >>
+            xcodebuild -downloadComponent MetalToolchain
-            brew install openmpi
+            mkdir -p ~/miniconda3
-            python<< parameters.python_version >> -m venv env
+            curl https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh -o ~/miniconda3/miniconda.sh
-            source env/bin/activate
+            bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
-            pip install --upgrade pip
+            rm ~/miniconda3/miniconda.sh
+            source ~/miniconda3/bin/activate
+            conda init --all
+            conda create -n env python=<< parameters.python_version >> -y
+            conda activate env
             pip install --upgrade cmake
             pip install nanobind==2.4.0
             pip install --upgrade setuptools
@@ -288,19 +311,19 @@ jobs:
       - run:
           name: Install Python package
           command: |
-            source env/bin/activate
+            conda activate env
             env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
               pip install . -v
       - run:
           name: Generate package stubs
           command: |
-            source env/bin/activate
+            conda activate env
             pip install typing_extensions
             python setup.py generate_stubs
       - run:
           name: Build Python package
           command: |
-            source env/bin/activate
+            conda activate env
             python setup.py clean --all
             << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
       - when:
@@ -310,7 +333,7 @@ jobs:
             - run:
                 name: Build common package
                 command: |
-                  source env/bin/activate
+                  conda activate env
                   python setup.py clean --all
                   << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
       - when:
@@ -319,7 +342,7 @@ jobs:
             - run:
                 name: Upload package
                 command: |
-                  source env/bin/activate
+                  conda activate env
                   twine upload dist/*
       - store_artifacts:
           path: dist/
@@ -392,7 +415,7 @@ jobs:
         default: ""
     machine:
       image: ubuntu-2204:current
-      resource_class: large
+      resource_class: xlarge
     steps:
       - checkout
       - run:
@@ -439,7 +462,7 @@ workflows:
       - mac_build_and_test:
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "14.0"]
+              macosx_deployment_target: ["13.5", "15.0"]
       - linux_build_and_test
       - cuda_build_and_test:
           matrix:
@@ -464,68 +487,7 @@ workflows:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["PYPI_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
+              xcode_version: ["26.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "PYPI_RELEASE=1"
       - build_documentation:
           filters:
             tags:
@@ -567,7 +529,7 @@ workflows:
           requires: [ hold ]
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "14.0"]
+              macosx_deployment_target: ["13.5", "15.0"]
       - linux_build_and_test:
           requires: [ hold ]
       - cuda_build_and_test:
@@ -586,53 +548,7 @@ workflows:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              xcode_version: ["16.2.0", "15.0.0"]
+              xcode_version: ["26.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
       - build_linux_release:
           matrix:
             parameters:
@@ -651,68 +567,7 @@ workflows:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
+              xcode_version: ["26.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
       - build_linux_release:
           matrix:
             parameters:

ACKNOWLEDGMENTS.md

@@ -19,12 +19,17 @@ MLX was developed with contributions from the following individuals:
 - Gleb Pobudzey: Added the `where` primitive, and groups in 1D and 2D convolutions.
 - Paul Paczuski: Improved stability of BCE loss calculation
 - Max-Heinrich Laves: Added `conv_transpose1d`, `conv_transpose2d`, and `conv_transpose3d` ops.
-- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer.
+- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer, and the `ReLU²` activation function.
 
 <a href="https://github.com/ml-explore/mlx/graphs/contributors">
   <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />
 </a>
 
+# Organizations
+
+MLX has received contributions from the following companies:
+- NVIDIA Corporation & Affiliates
+
 # Third-Party Software
 
 MLX leverages several third-party software, listed here together with

CMakeLists.txt

@@ -87,22 +87,21 @@ cmake_policy(SET CMP0135 NEW)
 
 add_library(mlx)
 
-if(MLX_BUILD_METAL)
-  set(METAL_LIB "-framework Metal")
-  set(FOUNDATION_LIB "-framework Foundation")
-  set(QUARTZ_LIB "-framework QuartzCore")
-endif()
-
 if(MLX_BUILD_CUDA)
   enable_language(CUDA)
 endif()
 
-if(MLX_BUILD_METAL AND NOT METAL_LIB)
+if(MLX_BUILD_METAL)
-  message(STATUS "Metal not found. Unable to build GPU")
+  find_library(METAL_LIB Metal)
-  set(MLX_BUILD_METAL OFF)
+  find_library(FOUNDATION_LIB Foundation)
-  set(MLX_METAL_DEBUG OFF)
+  find_library(QUARTZ_LIB QuartzCore)
-elseif(MLX_BUILD_METAL)
+  if(METAL_LIB)
-  message(STATUS "Building METAL sources")
+    message(STATUS "Metal found ${METAL_LIB}")
+  else()
+    message(
+      FATAL_ERROR
+        "Metal not found. Set MLX_BUILD_METAL=OFF to build without GPU")
+  endif()
 
   if(MLX_METAL_DEBUG)
     add_compile_definitions(MLX_METAL_DEBUG)
@@ -111,7 +110,8 @@ elseif(MLX_BUILD_METAL)
   # Throw an error if xcrun not found
   execute_process(
     COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
-    OUTPUT_VARIABLE MACOS_SDK_VERSION COMMAND_ERROR_IS_FATAL ANY)
+    OUTPUT_VARIABLE MACOS_SDK_VERSION
+    OUTPUT_STRIP_TRAILING_WHITESPACE COMMAND_ERROR_IS_FATAL ANY)
 
   if(${MACOS_SDK_VERSION} LESS 14.0)
     message(
@@ -140,6 +140,12 @@ elseif(MLX_BUILD_METAL)
   target_link_libraries(mlx PUBLIC ${METAL_LIB} ${FOUNDATION_LIB} ${QUARTZ_LIB})
 endif()
 
+if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
+  # With newer clang/gcc versions following libs are implicitly linked, but when
+  # building on old distributions they need to be explicitly listed.
+  target_link_libraries(mlx PRIVATE dl pthread)
+endif()
+
 if(WIN32)
   if(MSVC)
     # GGUF does not build with MSVC.

cmake/FindNCCL.cmake

@@ -0,0 +1,54 @@
+# FindNCCL.cmake This module finds the NVIDIA NCCL library and its include
+# directories.
+
+set(NCCL_ROOT_DIR
+    $ENV{NCCL_ROOT_DIR}
+    CACHE PATH "Folder contains NVIDIA NCCL")
+
+find_path(
+  NCCL_INCLUDE_DIRS
+  NAMES nccl.h
+  HINTS ${NCCL_INCLUDE_DIR} ${NCCL_ROOT_DIR} ${NCCL_ROOT_DIR}/include
+        ${CUDA_TOOLKIT_ROOT_DIR}/include)
+
+if($ENV{USE_STATIC_NCCL})
+  message(
+    STATUS "USE_STATIC_NCCL detected. Linking against static NCCL library")
+  set(NCCL_LIBNAME "libnccl_static.a")
+else()
+  set(NCCL_LIBNAME "nccl")
+endif()
+
+find_library(
+  NCCL_LIBRARIES
+  NAMES ${NCCL_LIBNAME}
+  HINTS ${NCCL_LIB_DIR}
+        ${NCCL_ROOT_DIR}
+        ${NCCL_ROOT_DIR}/lib
+        ${NCCL_ROOT_DIR}/lib/x86_64-linux-gnu
+        ${NCCL_ROOT_DIR}/lib64
+        ${CUDA_TOOLKIT_ROOT_DIR}/lib
+        ${CUDA_TOOLKIT_ROOT_DIR}/lib64)
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(NCCL DEFAULT_MSG NCCL_INCLUDE_DIRS
+                                  NCCL_LIBRARIES)
+
+if(NCCL_FOUND)
+  set(NCCL_HEADER_FILE "${NCCL_INCLUDE_DIRS}/nccl.h")
+  message(
+    STATUS "Determining NCCL version from the header file: ${NCCL_HEADER_FILE}")
+  file(
+    STRINGS ${NCCL_HEADER_FILE} NCCL_MAJOR_VERSION_DEFINED
+    REGEX "^[ \t]*#define[ \t]+NCCL_MAJOR[ \t]+[0-9]+.*$"
+    LIMIT_COUNT 1)
+  if(NCCL_MAJOR_VERSION_DEFINED)
+    string(REGEX REPLACE "^[ \t]*#define[ \t]+NCCL_MAJOR[ \t]+" ""
+                         NCCL_MAJOR_VERSION ${NCCL_MAJOR_VERSION_DEFINED})
+    message(STATUS "NCCL_MAJOR_VERSION: ${NCCL_MAJOR_VERSION}")
+  endif()
+  message(
+    STATUS
+      "Found NCCL (include: ${NCCL_INCLUDE_DIRS}, library: ${NCCL_LIBRARIES})")
+  mark_as_advanced(NCCL_ROOT_DIR NCCL_INCLUDE_DIRS NCCL_LIBRARIES)
+endif()

docs/requirements.txt

@@ -1,4 +1,5 @@
 sphinx
 breathe
 sphinx-book-theme
+sphinx-copybutton
 mlx

docs/src/conf.py

@@ -18,6 +18,7 @@ release = version
 # -- General configuration ---------------------------------------------------
 
 extensions = [
+    "sphinx_copybutton",
     "sphinx.ext.autodoc",
     "sphinx.ext.autosummary",
     "sphinx.ext.intersphinx",

docs/src/dev/custom_metal_kernels.rst

@@ -127,7 +127,8 @@ relying on a copy from ``ensure_row_contiguous``:
       name="myexp_strided",
       input_names=["inp"],
       output_names=["out"],
-      source=source
+      source=source,
+      ensure_row_contiguous=False,
   )
 
   def exp_elementwise(a: mx.array):
@@ -138,7 +139,6 @@ relying on a copy from ``ensure_row_contiguous``:
           threadgroup=(256, 1, 1),
           output_shapes=[a.shape],
           output_dtypes=[a.dtype],
-          ensure_row_contiguous=False,
       )
       return outputs[0]
 

docs/src/index.rst

@@ -70,6 +70,7 @@ are the CPU and GPU.
    python/fft
    python/linalg
    python/metal
+   python/cuda
    python/memory_management
    python/nn
    python/optimizers

docs/src/install.rst

@@ -271,7 +271,7 @@ and the CUDA toolkit. For example on Ubuntu, run the following:
    dpkg -i cuda-keyring_1.1-1_all.deb
    apt-get update -y
    apt-get -y install cuda-toolkit-12-9
-   apt-get install libblas-dev liblapack-dev liblapacke-dev -y
+   apt-get install libblas-dev liblapack-dev liblapacke-dev libcudnn9-dev-cuda-12 -y
 
 
 When building either the Python or C++ APIs make sure to pass the cmake flag

docs/src/python/cuda.rst

@@ -0,0 +1,9 @@
+CUDA
+=====
+
+.. currentmodule:: mlx.core.cuda
+
+.. autosummary::
+  :toctree: _autosummary
+
+  is_available

docs/src/python/fast.rst

@@ -13,3 +13,4 @@ Fast
   rope
   scaled_dot_product_attention
   metal_kernel
+  cuda_kernel

docs/src/python/nn/functions.rst

@@ -27,6 +27,7 @@ simple functions.
    mish
    prelu
    relu
+   relu2
    relu6
    selu
    sigmoid

docs/src/python/nn/layers.rst

@@ -50,6 +50,7 @@ Layers
    QuantizedLinear
    RMSNorm
    ReLU
+   ReLU2
    ReLU6
    RNN
    RoPE

docs/src/python/optimizers.rst

@@ -51,14 +51,14 @@ the saved state. Here's a simple example:
    optimizer.update(model, grads)
 
    # Save the state
-   state = tree_flatten(optimizer.state)
+   state = tree_flatten(optimizer.state, destination={})
-   mx.save_safetensors("optimizer.safetensors", dict(state))
+   mx.save_safetensors("optimizer.safetensors", state)
 
    # Later on, for example when loading from a checkpoint,
    # recreate the optimizer and load the state
    optimizer = optim.Adam(learning_rate=1e-2)
 
-   state = tree_unflatten(list(mx.load("optimizer.safetensors").items()))
+   state = tree_unflatten(mx.load("optimizer.safetensors"))
    optimizer.state = state
 
 Note, not every optimizer configuation parameter is saved in the state. For

docs/src/usage/compile.rst

@@ -130,8 +130,8 @@ Now make an array, and benchmark both functions:
 .. code-block:: python
 
   x = mx.random.uniform(shape=(32, 1000, 4096))
-  timeit(nn.gelu, x)
+  timeit(gelu, x)
-  timeit(mx.compile(nn.gelu), x)
+  timeit(mx.compile(gelu), x)
 
 On an M1 Max the times are 15.5 and 3.1 milliseconds. The compiled ``gelu`` is
 five times faster.
@@ -225,7 +225,7 @@ In some cases returning updated state can be pretty inconvenient. Hence,
   def fun(x, y):
       z = x + y
       state.append(z)
-      return mx.exp(z), state
+      return mx.exp(z)
 
   fun(mx.array(1.0), mx.array(2.0))
   # Prints [array(3, dtype=float32)]

docs/src/usage/distributed.rst

@@ -184,7 +184,7 @@ almost identical to the example above:
 
     def step(model, x, y):
         loss, grads = loss_grad_fn(model, x, y)
-        grads = mlx.nn.average_gradients(grads) # <---- This line was added
+        grads = mx.nn.average_gradients(grads)  # <---- This line was added
         optimizer.update(model, grads)
         return loss
 

docs/src/usage/export.rst

@@ -151,7 +151,7 @@ parameters, pass them as inputs to the ``call`` wrapper:
      model.update(tree_unflatten(list(params.items())))
      return model(x)
 
-   params = dict(tree_flatten(model.parameters()))
+   params = tree_flatten(model.parameters(), destination={})
    mx.export_function("model.mlxfn", call, (mx.zeros(4),), params)
 
 
@@ -164,11 +164,11 @@ to export a function which can be used for inputs with variable shapes:
 
 .. code-block:: python
 
-  mx.export_function("fun.mlxfn", mx.abs, mx.array(0.0), shapeless=True)
+  mx.export_function("fun.mlxfn", mx.abs, mx.array([0.0]), shapeless=True)
   imported_abs = mx.import_function("fun.mlxfn")
 
   # Ok
-  out, = imported_abs(mx.array(-1.0))
+  out, = imported_abs(mx.array([-1.0]))
 
   # Also ok
   out, = imported_abs(mx.array([-1.0, -2.0]))

docs/src/usage/indexing.rst

@@ -107,8 +107,20 @@ same array:
   >>> a
   array([1, 2, 0], dtype=int32)
 
+Note that unlike NumPy, slicing an array creates a copy, not a view. So
+mutating it does not mutate the original array:
 
-Note, unlike NumPy, updates to the same location are nondeterministic:
+.. code-block:: shell
+
+  >>> a = mx.array([1, 2, 3])
+  >>> b = a[:]
+  >>> b[2] = 0
+  >>> b
+  array([1, 2, 0], dtype=int32)
+  >>> a
+  array([1, 2, 3], dtype=int32)
+
+Also unlike NumPy, updates to the same location are nondeterministic:
 
 .. code-block:: shell
 

mlx/backend/common/utils.cpp

@@ -228,31 +228,4 @@ std::pair<Dims, Dims> get_grid_and_block_common(int dim0, int dim1, int dim2) {
       std::make_tuple(gx, gy, gz), std::make_tuple(bx, by, bz));
 }
 
-array swapaxes_in_eval(const array& x, int axis1, int axis2) {
-  int ndim = x.ndim();
-  if (axis1 < 0) {
-    axis1 += ndim;
-  }
-  if (axis2 < 0) {
-    axis2 += ndim;
-  }
-
-  auto shape = x.shape();
-  std::swap(shape[axis1], shape[axis2]);
-  auto strides = x.strides();
-  std::swap(strides[axis1], strides[axis2]);
-
-  auto [data_size, row_contiguous, col_contiguous] =
-      check_contiguity(shape, strides);
-  bool contiguous = data_size == x.data_size();
-
-  array out(std::move(shape), x.dtype(), nullptr, {});
-  out.copy_shared_buffer(
-      x,
-      std::move(strides),
-      {contiguous, row_contiguous, col_contiguous},
-      x.data_size());
-  return out;
-}
-
 } // namespace mlx::core

mlx/backend/common/utils.h

@@ -196,9 +196,6 @@ void shared_buffer_reshape(
     const Strides& out_strides,
     array& out);
 
-// Like the swapaxes op but safe to call in eval_gpu.
-array swapaxes_in_eval(const array& x, int axis1, int axis2);
-
 template <typename T>
 inline SmallVector<T> remove_index(SmallVector<T> vec, size_t index) {
   vec.erase(std::next(vec.begin(), index));

mlx/backend/cpu/compiled.cpp

@@ -15,6 +15,7 @@
 #include "mlx/backend/cpu/jit_compiler.h"
 #include "mlx/device.h"
 #include "mlx/graph_utils.h"
+#include "mlx/version.h"
 
 namespace mlx::core {
 
@@ -94,7 +95,11 @@ void* compile(
     kernel_file_name = kernel_name;
   }
 
-  auto output_dir = std::filesystem::temp_directory_path();
+  auto output_dir =
+      std::filesystem::temp_directory_path() / "mlx" / version() / "cpu";
+  if (!std::filesystem::exists(output_dir)) {
+    std::filesystem::create_directories(output_dir);
+  }
 
   std::string shared_lib_name = "lib" + kernel_file_name + ".so";
   auto shared_lib_path = (output_dir / shared_lib_name).string();
@@ -157,10 +162,12 @@ inline void build_kernel(
 #endif
 
   // Start the kernel
-  os << "void " << kernel_name << "(void** args) {" << std::endl;
+  os << "void " << kernel_name
+     << "(int* shape, int64_t** strides, void** args) {" << std::endl;
 
   // Add the input arguments
   int cnt = 0;
+  int strides_index = 1;
   for (size_t i = 0; i < inputs.size(); ++i) {
     // Skip constants from the input list
     if (is_constant(i)) {
@@ -175,8 +182,8 @@ inline void build_kernel(
        << "];" << std::endl;
     // Scalars and contiguous need no strides
     if (!is_scalar(x) && !contiguous) {
-      os << "  const size_t* " << xname << "_strides = (size_t*)args[" << cnt++
+      os << "  const int64_t* " << xname << "_strides = strides["
-         << "];" << std::endl;
+         << strides_index++ << "];" << std::endl;
     }
   }
 
@@ -186,10 +193,8 @@ inline void build_kernel(
     os << "  " << tstr << "* " << namer.get_name(x) << " = (" << tstr
        << "*)args[" << cnt++ << "];" << std::endl;
   }
-  // Add output strides and shape to extract the indices.
+  // Add output size
-  if (!contiguous) {
+  if (contiguous) {
-    os << "  const int* shape = (int*)args[" << cnt++ << "];" << std::endl;
-  } else {
     os << "  const size_t size = (size_t)args[" << cnt++ << "];" << std::endl;
   }
 
@@ -288,17 +293,8 @@ void Compiled::eval_cpu(
   auto [contiguous, shape, strides] =
       compiled_collapse_contiguous_dims(inputs, outputs[0], is_constant_);
 
-  // Force allocating shape/strides on heap so we can take their data() first
-  // and then std::move them.
-  // TODO: Refactor code to avoid heap allocation.
-  shape.grow();
-  for (auto& s : strides) {
-    s.grow();
-  }
-
   // Collect function input arguments.
   std::vector<void*> args;
-  int strides_index = 1;
   for (size_t i = 0; i < inputs.size(); ++i) {
     if (is_constant_(i)) {
       continue;
@@ -306,9 +302,6 @@ void Compiled::eval_cpu(
     const auto& x = inputs[i];
     encoder.set_input_array(x);
     args.push_back((void*)x.data<void>());
-    if (!contiguous && !is_scalar(x)) {
-      args.push_back(strides[strides_index++].data());
-    }
   }
 
   // Get the kernel name from the lib
@@ -343,16 +336,20 @@ void Compiled::eval_cpu(
     args.push_back(x.data<void>());
     encoder.set_output_array(x);
   }
-  if (!contiguous) {
+  if (contiguous) {
-    args.push_back((void*)shape.data());
-  } else {
     args.push_back((void*)outputs[0].data_size());
   }
-  auto fun = (void (*)(void**))fn_ptr;
+  auto fun = reinterpret_cast<void (*)(int*, int64_t**, void**)>(fn_ptr);
   encoder.dispatch([fun,
                     args = std::move(args),
                     strides = std::move(strides),
-                    shape = std::move(shape)]() mutable { fun(args.data()); });
+                    shape = std::move(shape)]() mutable {
+    SmallVector<int64_t*> strides_ptrs;
+    for (auto& s : strides) {
+      strides_ptrs.push_back(s.data());
+    }
+    fun(shape.data(), strides_ptrs.data(), args.data());
+  });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/lapack.h

@@ -47,7 +47,7 @@ INSTANTIATE_LAPACK_REAL(orgqr)
 INSTANTIATE_LAPACK_REAL(syevd)
 INSTANTIATE_LAPACK_REAL(geev)
 INSTANTIATE_LAPACK_REAL(potrf)
-INSTANTIATE_LAPACK_REAL(gesvdx)
+INSTANTIATE_LAPACK_REAL(gesdd)
 INSTANTIATE_LAPACK_REAL(getrf)
 INSTANTIATE_LAPACK_REAL(getri)
 INSTANTIATE_LAPACK_REAL(trtri)

mlx/backend/cpu/quantized.cpp

@@ -1,7 +1,5 @@
 // Copyright © 2023 Apple Inc.
 
-#include <cassert>
-
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
@@ -13,6 +11,35 @@ namespace mlx::core {
 
 namespace {
 
+const static float MXFP4_LUT[16] = {
+    +0.0f,
+    +0.5f,
+    +1.0f,
+    +1.5f,
+    +2.0f,
+    +3.0f,
+    +4.0f,
+    +6.0f,
+    -0.0f,
+    -0.5f,
+    -1.0f,
+    -1.5f,
+    -2.0f,
+    -3.0f,
+    -4.0f,
+    -6.0f};
+
+template <typename T>
+static inline T dequantize_scale(uint8_t s) {
+  using FOrI = union {
+    bfloat16_t f;
+    uint16_t i;
+  };
+  FOrI out;
+  out.i = (s == 0 ? 0x40 : (static_cast<uint16_t>(s) << 7));
+  return static_cast<T>(out.f);
+}
+
 inline constexpr short get_pack_factor(int bits, int wsize = 8) {
   return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
 }
@@ -407,6 +434,231 @@ void _qmm_dispatch(
   }
 }
 
+template <typename T>
+void mxfp4_qmm(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = get_pack_factor(4, 8);
+  constexpr int bytes_per_pack = get_bytes_per_pack(4);
+  constexpr int packs_in_group = group_size / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint8_t* w_local = (const uint8_t*)w;
+    const uint8_t* scales_local = scales;
+
+    std::fill(result, result + N, 0);
+
+    for (int k = 0; k < K; k++) {
+      T* result_local = result;
+      T xi = *x++;
+
+      for (int n = 0; n < N; n += group_size) {
+        T scale = dequantize_scale<T>(*scales_local++);
+        for (int ng = 0; ng < packs_in_group; ng++) {
+          uint8_t wi = *w_local++;
+#pragma clang loop unroll(full)
+          for (int p = 0; p < pack_factor; p++) {
+            (*result_local++) +=
+                xi * scale * static_cast<T>(MXFP4_LUT[wi & 0xf]);
+            wi >>= 4;
+          }
+        }
+      }
+    }
+
+    result += N;
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_t(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = get_pack_factor(4, 8);
+  constexpr int bytes_per_pack = get_bytes_per_pack(4);
+  constexpr int packs_in_group = group_size / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint8_t* w_local = (const uint8_t*)w;
+    const uint8_t* scales_local = scales;
+
+    for (int n = 0; n < N; n++) {
+      const T* x_local = x;
+      T sum = 0;
+      for (int k = 0; k < K; k += group_size) {
+        T scale = dequantize_scale<T>(*scales_local++);
+
+        T gsum = 0;
+        for (int kw = 0; kw < packs_in_group; kw++) {
+          uint8_t wi = *w_local++;
+#pragma clang loop unroll(full)
+          for (int p = 0; p < pack_factor; p++) {
+            gsum += (*x_local++) * static_cast<T>(MXFP4_LUT[wi & 0xf]);
+            wi >>= 4;
+          }
+        }
+        sum += scale * gsum;
+      }
+      *result = sum;
+      result++;
+    }
+
+    x += K;
+  }
+}
+
+template <int S>
+simd::Simd<float, S> mxfp4_extract_bits_simd(const uint32_t* w) {
+  if constexpr (S == 8) {
+    constexpr std::array<uint32_t, 8> shifts_ = {{0, 4, 8, 12, 16, 20, 24, 28}};
+    auto shifts(*(simd::Simd<uint32_t, S>*)&shifts_);
+    auto wi = simd::Simd<uint32_t, S>(*w);
+    wi = wi >> shifts;
+    wi = wi & 0xf;
+    simd::Simd<float, S> w_out;
+    for (int i = 0; i < S; ++i) {
+      w_out[i] = MXFP4_LUT[wi[i]];
+    }
+    return w_out;
+  } else {
+    // Appease compiler.. but should never get here
+    throw std::runtime_error("Unsupported combination for simd qmm.");
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_t_simd(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = 32 / 4;
+  constexpr int packs_in_group = group_size / pack_factor;
+  constexpr int S = simd::max_size<T>;
+  static_assert(
+      S % pack_factor == 0, "SIMD size must be divisible by pack factor");
+  constexpr int packs_per_simd = S / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint32_t* w_local = w;
+    const uint8_t* scales_local = scales;
+
+    for (int n = 0; n < N; n++) {
+      simd::Simd<float, S> acc(0);
+      auto x_local = x;
+      for (int k = 0; k < K; k += group_size) {
+        T scale = dequantize_scale<T>(*scales_local++);
+
+        simd::Simd<float, S> g_acc(0);
+        for (int kw = 0; kw < packs_in_group; kw += packs_per_simd) {
+          // Extract bits
+          auto wf = mxfp4_extract_bits_simd<S>(w_local);
+          w_local += packs_per_simd;
+          simd::Simd<float, S> x_simd = simd::load<T, S>(x_local);
+          g_acc = g_acc + x_simd * wf;
+          x_local += S;
+        }
+        acc = acc + scale * g_acc;
+      }
+
+      *result = T(simd::sum(acc));
+      result++;
+    }
+    x += K;
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_dispatch_transpose(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K,
+    bool transposed_w) {
+  if (transposed_w) {
+    // the simd size must be a multiple of the number of elements per word
+    if constexpr (simd::max_size<T> % 8 == 0) {
+      mxfp4_qmm_t_simd<T>(result, x, w, scales, M, N, K);
+    } else {
+      mxfp4_qmm_t<T>(result, x, w, scales, M, N, K);
+    }
+  } else {
+    mxfp4_qmm<T>(result, x, w, scales, M, N, K);
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_dispatch_typed(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    bool transposed_w) {
+  int K = x.shape(-1);
+  int M = x.ndim() > 1 ? x.shape(-2) : 1;
+  int N = out.shape(-1);
+  int w_els = w.ndim() > 2 ? w.shape(-1) * w.shape(-2) : 0;
+  int g_els = w.ndim() > 2 ? scales.shape(-1) * scales.shape(-2) : 0;
+  int batch_size = x.size() / (K * M);
+
+  auto out_ptr = out.data<T>();
+  auto x_ptr = x.data<T>();
+  auto w_ptr = w.data<uint32_t>();
+  auto scales_ptr = scales.data<uint8_t>();
+  for (int i = 0; i < batch_size; i++) {
+    mxfp4_qmm_dispatch_transpose<T>(
+        out_ptr + i * M * N,
+        x_ptr + elem_to_loc(i * M * K, x.shape(), x.strides()),
+        w_ptr + elem_to_loc(i * w_els, w.shape(), w.strides()),
+        scales_ptr + elem_to_loc(i * g_els, scales.shape(), scales.strides()),
+        M,
+        N,
+        K,
+        transposed_w);
+  }
+}
+
+void mxfp4_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    bool transposed_w) {
+  switch (x.dtype()) {
+    case bfloat16:
+      mxfp4_qmm_dispatch_typed<bfloat16_t>(out, x, w, scales, transposed_w);
+      break;
+    case float16:
+      mxfp4_qmm_dispatch_typed<float16_t>(out, x, w, scales, transposed_w);
+      break;
+    case float32:
+      mxfp4_qmm_dispatch_typed<float>(out, x, w, scales, transposed_w);
+      break;
+    default:
+      throw std::invalid_argument(
+          "[quantized_matmul] only floating types are supported");
+  }
+}
+
 template <typename T>
 void _bs_qmm_dispatch_typed(
     array& out,
@@ -513,41 +765,106 @@ void _bs_qmm_dispatch(
   }
 }
 
+template <typename T>
+void mxfp4_bs_qmm_dispatch_typed(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    bool transposed_w) {
+  int K = x.shape(-1);
+  int M = x.shape(-2);
+  int N = out.shape(-1);
+
+  int w_els = w.shape(-1) * w.shape(-2);
+  int g_els = scales.shape(-1) * scales.shape(-2);
+
+  auto out_ptr = out.data<T>();
+  auto x_ptr = x.data<T>();
+  auto w_ptr = w.data<uint32_t>();
+  auto scales_ptr = scales.data<uint8_t>();
+  auto lhs_indices_ptr = lhs_indices.data<uint32_t>();
+  auto rhs_indices_ptr = rhs_indices.data<uint32_t>();
+
+  for (int i = 0; i < lhs_indices.size(); i++) {
+    int x_idx = lhs_indices_ptr[elem_to_loc(
+        i, lhs_indices.shape(), lhs_indices.strides())];
+    int w_idx = rhs_indices_ptr[elem_to_loc(
+        i, rhs_indices.shape(), rhs_indices.strides())];
+    mxfp4_qmm_dispatch_transpose<T>(
+        out_ptr + i * M * N,
+        x_ptr + elem_to_loc(x_idx * M * K, x.shape(), x.strides()),
+        w_ptr + elem_to_loc(w_idx * w_els, w.shape(), w.strides()),
+        scales_ptr +
+            elem_to_loc(w_idx * g_els, scales.shape(), scales.strides()),
+        M,
+        N,
+        K,
+        transposed_w);
+  }
+}
+
+void mxfp4_bs_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    bool transposed_w) {
+  switch (x.dtype()) {
+    case float32:
+      mxfp4_bs_qmm_dispatch_typed<float>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    case float16:
+      mxfp4_bs_qmm_dispatch_typed<float16_t>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    case bfloat16:
+      mxfp4_bs_qmm_dispatch_typed<bfloat16_t>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    default:
+      throw std::invalid_argument(
+          "[quantized_matmul] only floating types are supported");
+  }
+}
+
 } // namespace
 
 void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 4);
-
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
-  auto& biases_pre = inputs[3];
 
-  std::vector<array> temps;
+  auto& encoder = cpu::get_command_encoder(stream());
-  auto ensure_row_contiguous = [s = stream(), &temps](const array& arr) {
+  auto ensure_row_contiguous = [s = stream(), &encoder](const array& arr) {
     if (arr.flags().row_contiguous) {
       return arr;
     } else {
-      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
+      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
-      copy_cpu(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, arr_cpy, CopyType::General, s);
-      return temps.back();
+      encoder.add_temporary(arr_cpy);
+      return arr_cpy;
     }
   };
 
   auto x = ensure_row_contiguous(x_pre);
   auto w = ensure_row_contiguous(w_pre);
   auto scales = ensure_row_contiguous(scales_pre);
-  auto biases = ensure_row_contiguous(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
-  auto& encoder = cpu::get_command_encoder(stream());
-  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
-  encoder.set_input_array(biases);
   encoder.set_output_array(out);
+  if (mode_ == QuantizationMode::Affine) {
+    auto biases = ensure_row_contiguous(inputs[3]);
+    encoder.set_input_array(biases);
     encoder.dispatch([out = array::unsafe_weak_copy(out),
                       x = array::unsafe_weak_copy(x),
                       w = array::unsafe_weak_copy(w),
@@ -558,48 +875,54 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
                       transpose_ = transpose_]() mutable {
       _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
     });
+  } else {
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      transpose_ = transpose_]() mutable {
+      mxfp4_qmm_dispatch(out, x, w, scales, transpose_);
+    });
+  }
 }
 
 void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 6);
-
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
-  auto& biases_pre = inputs[3];
+  auto& lhs_indices = inputs[inputs.size() - 2];
-  auto& lhs_indices = inputs[4];
+  auto& rhs_indices = inputs[inputs.size() - 1];
-  auto& rhs_indices = inputs[5];
 
-  std::vector<array> temps;
+  auto& encoder = cpu::get_command_encoder(stream());
   auto ensure_row_contiguous_last_dims = [s = stream(),
-                                          &temps](const array& arr) {
+                                          &encoder](const array& arr) {
     auto stride_0 = arr.strides()[arr.ndim() - 2];
     auto stride_1 = arr.strides()[arr.ndim() - 1];
     if (stride_0 == arr.shape(-1) && stride_1 == 1) {
       return arr;
     } else {
-      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
+      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
-      copy_cpu(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, arr_cpy, CopyType::General, s);
-      return temps.back();
+      encoder.add_temporary(arr_cpy);
+      return arr_cpy;
     }
   };
 
   auto x = ensure_row_contiguous_last_dims(x_pre);
   auto w = ensure_row_contiguous_last_dims(w_pre);
   auto scales = ensure_row_contiguous_last_dims(scales_pre);
-  auto biases = ensure_row_contiguous_last_dims(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
-  auto& encoder = cpu::get_command_encoder(stream());
-  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
-  encoder.set_input_array(biases);
   encoder.set_input_array(lhs_indices);
   encoder.set_input_array(rhs_indices);
   encoder.set_output_array(out);
+  if (mode_ == QuantizationMode::Affine) {
+    auto biases = ensure_row_contiguous_last_dims(inputs[3]);
+    encoder.set_input_array(biases);
     encoder.dispatch([out = array::unsafe_weak_copy(out),
                       x = array::unsafe_weak_copy(x),
                       w = array::unsafe_weak_copy(w),
@@ -622,6 +945,18 @@ void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
           bits_,
           transpose_);
     });
+  } else {
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      lhs_indices = array::unsafe_weak_copy(lhs_indices),
+                      rhs_indices = array::unsafe_weak_copy(rhs_indices),
+                      transpose_ = transpose_]() mutable {
+      mxfp4_bs_qmm_dispatch(
+          out, x, w, scales, lhs_indices, rhs_indices, transpose_);
+    });
+  }
 }
 
 template <typename T, typename U>
@@ -705,7 +1040,7 @@ void dispatch_quantize(
       w_ptr, out_ptr, scales_ptr, biases_ptr, bits, group_size, w.size());
 }
 
-void fast::AffineQuantize::eval_cpu(
+void fast::Quantize::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
   auto ensure_row_contiguous = [s = stream()](const array& arr) {
@@ -764,7 +1099,7 @@ void fast::AffineQuantize::eval_cpu(
       }
     } else {
       throw std::runtime_error(
-          "[fast::AffineQuantize::eval_cpu] Only supports floating point inputs");
+          "[fast::Quantize::eval_cpu] Only supports floating point inputs");
     }
   });
 }

mlx/backend/cpu/reduce.cpp

@@ -491,19 +491,27 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
         switch (in.dtype()) {
           case bool_:
           case uint8:
+            reduce_dispatch_sum_prod<uint8_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint16:
+            reduce_dispatch_sum_prod<uint16_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint32:
+            reduce_dispatch_sum_prod<uint32_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint64:
+            reduce_dispatch_sum_prod<uint64_t>(in, out, reduce_type_, axes_);
+            break;
           case int8:
             reduce_dispatch_sum_prod<int8_t>(in, out, reduce_type_, axes_);
             break;
           case int16:
-          case uint16:
             reduce_dispatch_sum_prod<int16_t>(in, out, reduce_type_, axes_);
             break;
           case int32:
-          case uint32:
             reduce_dispatch_sum_prod<int32_t>(in, out, reduce_type_, axes_);
             break;
           case int64:
-          case uint64:
             reduce_dispatch_sum_prod<int64_t>(in, out, reduce_type_, axes_);
             break;
           case float16:

mlx/backend/cpu/simd/accelerate_simd.h

@@ -234,6 +234,7 @@ Simd<T, N> remainder(Simd<T, N> a, Simd<T, N> b) {
 
 template <typename MaskT, typename T1, typename T2, int N>
 Simd<T1, N> select(Simd<MaskT, N> mask, Simd<T1, N> x, Simd<T2, N> y) {
+  static_assert(std::is_same_v<MaskT, bool>);
   if constexpr (sizeof(T1) == 1) {
     return asd::bitselect(y.value, x.value, asd::convert<char>(mask.value));
   } else if constexpr (sizeof(T1) == 2) {
@@ -251,9 +252,13 @@ Simd<T, N> pow(Simd<T, N> base, Simd<T, N> exp) {
     return asd::pow(base.value, exp.value);
   } else {
     Simd<T, N> res = 1;
-    while (any(exp)) {
+    // Raising an integer to a negative power is undefined
-      res = select(exp & 1, res * base, res);
+    if (any(exp < 0)) {
-      base = select(exp, base * base, base);
+      return 0;
+    }
+    while (any(exp > 0)) {
+      res = select((exp & 1) != 0, res * base, res);
+      base = select(exp > 0, base * base, base);
       exp = exp >> 1;
     }
     return res;

mlx/backend/cpu/svd.cpp

@@ -81,9 +81,7 @@ void svd_impl(
     // Vᵀ of shape N x N. (M x M in lapack).
     const int ldvt = M;
 
-    auto job_u = (u_ptr) ? "V" : "N";
+    auto jobz = (u_ptr) ? "A" : "N";
-    auto job_vt = (u_ptr) ? "V" : "N";
-    static constexpr auto range = "A";
 
     // Will contain the number of singular values after the call has returned.
     int ns = 0;
@@ -91,30 +89,20 @@ void svd_impl(
 
     // Will contain the indices of eigenvectors that failed to converge (not
     // used here but required by lapack).
-    auto iwork = array::Data{allocator::malloc(sizeof(int) * 12 * K)};
+    auto iwork = array::Data{allocator::malloc(sizeof(int) * 8 * K)};
 
     static const int lwork_query = -1;
 
-    static const int ignored_int = 0;
-    static const T ignored_float = 0;
-
     int info;
 
     // Compute workspace size.
-    gesvdx<T>(
+    gesdd<T>(
-        /* jobu = */ job_u,
+        /* jobz = */ jobz,
-        /* jobvt = */ job_vt,
-        /* range = */ range,
         // M and N are swapped since lapack expects column-major.
         /* m = */ &N,
         /* n = */ &M,
         /* a = */ nullptr,
         /* lda = */ &lda,
-        /* vl = */ &ignored_float,
-        /* vu = */ &ignored_float,
-        /* il = */ &ignored_int,
-        /* iu = */ &ignored_int,
-        /* ns = */ &ns,
         /* s = */ nullptr,
         /* u = */ nullptr,
         /* ldu = */ &ldu,
@@ -136,20 +124,13 @@ void svd_impl(
 
     // Loop over matrices.
     for (int i = 0; i < num_matrices; i++) {
-      gesvdx<T>(
+      gesdd<T>(
-          /* jobu = */ job_u,
+          /* jobz = */ jobz,
-          /* jobvt = */ job_vt,
-          /* range = */ range,
           // M and N are swapped since lapack expects column-major.
           /* m = */ &N,
           /* n = */ &M,
           /* a = */ in_ptr + M * N * i,
           /* lda = */ &lda,
-          /* vl = */ &ignored_float,
-          /* vu = */ &ignored_float,
-          /* il = */ &ignored_int,
-          /* iu = */ &ignored_int,
-          /* ns = */ &ns,
           /* s = */ s_ptr + K * i,
           // According to the identity above, lapack will write Vᵀᵀ as U.
           /* u = */ vt_ptr ? vt_ptr + N * N * i : nullptr,
@@ -167,13 +148,6 @@ void svd_impl(
         ss << "svd_impl: sgesvdx_ failed with code " << info;
         throw std::runtime_error(ss.str());
       }
-
-      if (ns != K) {
-        std::stringstream ss;
-        ss << "svd_impl: expected " << K << " singular values, but " << ns
-           << " were computed.";
-        throw std::runtime_error(ss.str());
-      }
     }
   });
   encoder.add_temporary(in);

mlx/backend/cuda/CMakeLists.txt

@@ -8,7 +8,6 @@ target_sources(
   PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/arange.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/binary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/binary_two.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/copy.cu
@@ -17,8 +16,13 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general_dynamic.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general_input.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/conv/gemm_conv.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/conv/gemm_grouped_conv.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/cuda.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/cudnn_utils.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/custom_kernel.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/distributed.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/eval.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/event.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/fence.cpp
@@ -45,18 +49,20 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/sort.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/ternary.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/unary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/quantized/affine_quantize.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/quantized/quantized.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/worker.cpp)
 
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/binary)
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/unary)
+
 if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.9.0)
   target_sources(
-    mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_batched_gemm_12_9.cu)
+    mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm_batched_12_9.cu)
 else()
   target_sources(
-    mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_batched_gemm_12_0.cpp)
+    mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm_batched_12_0.cpp)
 endif()
 
 target_compile_definitions(mlx PRIVATE MLX_USE_CUDA)
@@ -148,7 +154,7 @@ target_link_libraries(mlx PRIVATE CUDA::nvrtc CUDA::cuda_driver)
 FetchContent_Declare(
   cudnn
   GIT_REPOSITORY https://github.com/NVIDIA/cudnn-frontend.git
-  GIT_TAG v1.12.1
+  GIT_TAG v1.14.0
   GIT_SHALLOW TRUE
   EXCLUDE_FROM_ALL)
 set(CUDNN_FRONTEND_SKIP_JSON_LIB ON)

mlx/backend/cuda/allocator.cpp

@@ -30,8 +30,15 @@ SmallSizePool::SmallSizePool() {
   next_free_ = buffer_;
 
   CHECK_CUDA_ERROR(cudaMallocManaged(&data_, small_pool_size));
+#if CUDART_VERSION >= 13000
+  cudaMemLocation loc;
+  loc.type = cudaMemLocationTypeDevice;
+  loc.id = 0;
+#else
+  int loc = 0;
+#endif // CUDART_VERSION >= 13000
   CHECK_CUDA_ERROR(
-      cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetReadMostly, 0));
+      cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetReadMostly, loc));
 
   auto curr = next_free_;
   for (size_t i = 1; i < num_blocks; ++i) {

mlx/backend/cuda/arange.cu

@@ -6,23 +6,33 @@
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
 
+#include <cooperative_groups.h>
 #include <nvtx3/nvtx3.hpp>
-#include <thrust/device_ptr.h>
-#include <thrust/transform.h>
 
 namespace mlx::core {
 
 namespace cu {
 
-template <typename T>
+namespace cg = cooperative_groups;
-struct Arange {
-  const T start;
-  const T step;
 
-  __device__ T operator()(uint32_t i) const {
+template <typename T, typename IdxT, int N_WRITES>
-    return start + i * step;
+__global__ void arange(T* out, IdxT size, T start, T step) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_WRITES > size) {
+    for (IdxT i = index * N_WRITES; i < size; ++i) {
+      out[i] = start + i * step;
     }
-};
+  } else {
+    AlignedVector<T, N_WRITES> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_WRITES; ++i) {
+      out_vec[i] = start + (index * N_WRITES + i) * step;
+    }
+
+    store_vector<N_WRITES>(out, index, out_vec);
+  }
+}
 
 } // namespace cu
 
@@ -36,19 +46,23 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& encoder = cu::get_command_encoder(stream());
   encoder.set_output_array(out);
 
-  auto capture = encoder.capture_context();
   dispatch_int_float_types(out.dtype(), "Arange", [&](auto type_tag) {
     using CTYPE = MLX_GET_TYPE(type_tag);
     using OutType = cuda_type_t<CTYPE>;
-    CTYPE step =
+    constexpr int N_WRITES = 16 / sizeof(OutType);
-        static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_);
+    dispatch_bool(out.data_size() > INT32_MAX, [&](auto large) {
-    thrust::transform(
+      using IdxT = std::conditional_t<large(), int64_t, int32_t>;
-        cu::thrust_policy(encoder.stream()),
+      auto [num_blocks, block_dims] = get_launch_args(out, large(), N_WRITES);
-        thrust::counting_iterator<uint32_t>(0),
+      encoder.add_kernel_node(
-        thrust::counting_iterator<uint32_t>(out.data_size()),
+          cu::arange<OutType, IdxT, N_WRITES>,
-        thrust::device_pointer_cast(out.data<OutType>()),
+          num_blocks,
-        cu::Arange<OutType>{
+          block_dims,
-            static_cast<OutType>(start_), static_cast<OutType>(step)});
+          0,
+          out.data<OutType>(),
+          out.data_size(),
+          static_cast<CTYPE>(start_),
+          static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_));
+    });
   });
 }
 

mlx/backend/cuda/binary/CMakeLists.txt

@@ -0,0 +1,21 @@
+target_sources(
+  mlx
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/add.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arctan2.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/bitwise_binary.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/divide.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/equal.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/greater.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/greater_equal.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/less.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/less_equal.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/logical_and.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/logical_or.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/log_add_exp.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/minimum.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/maximum.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/multiply.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/power.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/remainder.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/not_equal.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/subtract.cu)

mlx/backend/cuda/binary/add.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(Add)
+} // namespace mlx::core

mlx/backend/cuda/binary/arctan2.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(ArcTan2)
+} // namespace mlx::core

mlx/backend/cuda/binary/binary.cuh

@@ -99,39 +99,89 @@ __global__ void binary_vv(const In* a, const In* b, Out* out, IdxT size) {
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
+template <
+    typename Op,
+    typename In,
+    typename Out,
+    typename IdxT,
+    int NDIM,
+    int N_READS>
 __global__ void binary_g_nd(
     const In* a,
     const In* b,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
-  IdxT index = cg::this_grid().thread_rank();
+  auto block = cg::this_thread_block();
-  if (index < size) {
+  auto grid = cg::this_grid();
-    auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
+  IdxT index_rest =
-        index, shape.data(), a_strides.data(), b_strides.data());
+      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-    out[index] = Op{}(a[a_idx], b[b_idx]);
+  if (index_rest >= size_rest) {
+    return;
   }
+
+  auto shape_x = shape[NDIM - 1];
+  auto a_stride_x = a_strides[NDIM - 1];
+  auto b_stride_x = b_strides[NDIM - 1];
+  IdxT index_x =
+      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
+      index_rest * shape_x, shape.data(), a_strides.data(), b_strides.data());
+  auto a_vec =
+      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, In(0));
+  auto b_vec =
+      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, In(0));
+
+  AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    out_vec[i] = Op{}(a_vec[i], b_vec[i]);
+  }
+  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void binary_g(
     const In* a,
     const In* b,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides a_strides,
     const __grid_constant__ Strides b_strides,
     int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
+  auto block = cg::this_thread_block();
-  if (index < size) {
+  auto grid = cg::this_grid();
-    auto [a_idx, b_idx] = elem_to_loc(
+  IdxT index_rest =
-        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
+      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-    out[index] = Op{}(a[a_idx], b[b_idx]);
+  if (index_rest >= size_rest) {
+    return;
   }
+
+  auto shape_x = shape[ndim - 1];
+  auto a_stride_x = a_strides[ndim - 1];
+  auto b_stride_x = b_strides[ndim - 1];
+  IdxT index_x =
+      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  auto [a_idx, b_idx] = elem_to_loc(
+      index_rest * shape_x,
+      shape.data(),
+      a_strides.data(),
+      b_strides.data(),
+      ndim);
+  auto a_vec =
+      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, In(0));
+  auto b_vec =
+      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, In(0));
+
+  AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    out_vec[i] = Op{}(a_vec[i], b_vec[i]);
+  }
+  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
 template <typename Op, typename In, typename Out>
@@ -209,39 +259,61 @@ void binary_op_gpu_inplace(
                 auto& a_strides = strides[0];
                 auto& b_strides = strides[1];
                 int ndim = shape.size();
+                int work_per_thread = 1;
+                auto dim0 = ndim > 0 ? shape.back() : 1;
+                auto rest = out.size() / dim0;
+                if (dim0 >= 4) {
+                  work_per_thread = 4;
+                }
+                dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
+                auto block_dims = get_block_dims(dim0, rest, 1);
+                uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
+                uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
                 if (ndim <= 3) {
                   dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                    auto [num_blocks, block_dims] =
+                    auto kernel = cu::binary_g_nd<
-                        get_launch_args(out, large());
-                    encoder.add_kernel_node(
-                        cu::binary_g_nd<
                         Op,
                         InType,
                         OutType,
                         IdxT,
-                            dims_constant()>,
+                        dims_constant(),
-                        num_blocks,
+                        1>;
+                    if (work_per_thread == 4) {
+                      kernel = cu::binary_g_nd<
+                          Op,
+                          InType,
+                          OutType,
+                          IdxT,
+                          dims_constant(),
+                          4>;
+                    }
+                    encoder.add_kernel_node(
+                        kernel,
+                        {num_blocks_x, num_blocks_y},
                         block_dims,
                         0,
                         a.data<InType>(),
                         b.data<InType>(),
                         out.data<OutType>(),
-                        out.size(),
+                        rest,
                         const_param<dims_constant()>(shape),
                         const_param<dims_constant()>(a_strides),
                         const_param<dims_constant()>(b_strides));
                   });
                 } else {
-                  auto [num_blocks, block_dims] = get_launch_args(out, large());
+                  auto kernel = cu::binary_g<Op, InType, OutType, IdxT, 1>;
+                  if (work_per_thread == 4) {
+                    kernel = cu::binary_g<Op, InType, OutType, IdxT, 4>;
+                  }
                   encoder.add_kernel_node(
-                      cu::binary_g<Op, InType, OutType, IdxT>,
+                      kernel,
-                      num_blocks,
+                      {num_blocks_x, num_blocks_y},
                       block_dims,
                       0,
                       a.data<InType>(),
                       b.data<InType>(),
                       out.data<OutType>(),
-                      out.size(),
+                      rest,
                       const_param(shape),
                       const_param(a_strides),
                       const_param(b_strides),
@@ -304,54 +376,4 @@ void binary_op_gpu(
     binary_op_gpu<cu::func>(inputs, out, name(), s);                  \
   }
 
-BINARY_GPU(Add)
-BINARY_GPU(ArcTan2)
-BINARY_GPU(Divide)
-BINARY_GPU(Remainder)
-BINARY_GPU(Greater)
-BINARY_GPU(GreaterEqual)
-BINARY_GPU(Less)
-BINARY_GPU(LessEqual)
-BINARY_GPU(LogicalAnd)
-BINARY_GPU(LogicalOr)
-BINARY_GPU(LogAddExp)
-BINARY_GPU(Maximum)
-BINARY_GPU(Minimum)
-BINARY_GPU(Multiply)
-BINARY_GPU(NotEqual)
-BINARY_GPU(Power)
-BINARY_GPU(Subtract)
-
-void Equal::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("Equal::eval_gpu");
-  auto& s = out.primitive().stream();
-  if (equal_nan_) {
-    binary_op_gpu<cu::NaNEqual>(inputs, out, name(), s);
-  } else {
-    binary_op_gpu<cu::Equal>(inputs, out, name(), s);
-  }
-}
-
-void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("BitwiseBinary::eval_gpu");
-  auto& s = out.primitive().stream();
-  switch (op_) {
-    case BitwiseBinary::And:
-      binary_op_gpu<cu::BitwiseAnd>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::Or:
-      binary_op_gpu<cu::BitwiseOr>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::Xor:
-      binary_op_gpu<cu::BitwiseXor>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::LeftShift:
-      binary_op_gpu<cu::LeftShift>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::RightShift:
-      binary_op_gpu<cu::RightShift>(inputs, out, name(), s);
-      break;
-  }
-}
-
 } // namespace mlx::core

mlx/backend/cuda/binary/bitwise_binary.cu

@@ -0,0 +1,27 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("BitwiseBinary::eval_gpu");
+  auto& s = out.primitive().stream();
+  switch (op_) {
+    case BitwiseBinary::And:
+      binary_op_gpu<cu::BitwiseAnd>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::Or:
+      binary_op_gpu<cu::BitwiseOr>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::Xor:
+      binary_op_gpu<cu::BitwiseXor>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::LeftShift:
+      binary_op_gpu<cu::LeftShift>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::RightShift:
+      binary_op_gpu<cu::RightShift>(inputs, out, name(), s);
+      break;
+  }
+}
+} // namespace mlx::core

mlx/backend/cuda/binary/divide.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(Divide)
+} // namespace mlx::core

mlx/backend/cuda/binary/equal.cu

@@ -0,0 +1,15 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+void Equal::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Equal::eval_gpu");
+  auto& s = out.primitive().stream();
+  if (equal_nan_) {
+    binary_op_gpu<cu::NaNEqual>(inputs, out, name(), s);
+  } else {
+    binary_op_gpu<cu::Equal>(inputs, out, name(), s);
+  }
+}
+} // namespace mlx::core

mlx/backend/cuda/binary/greater.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(Greater)
+} // namespace mlx::core

mlx/backend/cuda/binary/greater_equal.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(GreaterEqual)
+} // namespace mlx::core

mlx/backend/cuda/binary/less.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(Less)
+} // namespace mlx::core

mlx/backend/cuda/binary/less_equal.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(LessEqual)
+} // namespace mlx::core

mlx/backend/cuda/binary/log_add_exp.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(LogAddExp)
+} // namespace mlx::core

mlx/backend/cuda/binary/logical_and.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(LogicalAnd)
+} // namespace mlx::core

mlx/backend/cuda/binary/logical_or.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(LogicalOr)
+} // namespace mlx::core

mlx/backend/cuda/binary/maximum.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(Maximum)
+} // namespace mlx::core

mlx/backend/cuda/binary/minimum.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(Minimum)
+} // namespace mlx::core

mlx/backend/cuda/binary/multiply.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(Multiply)
+} // namespace mlx::core

mlx/backend/cuda/binary/not_equal.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(NotEqual)
+} // namespace mlx::core

mlx/backend/cuda/binary/power.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(Power)
+} // namespace mlx::core

mlx/backend/cuda/binary/remainder.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(Remainder)
+} // namespace mlx::core

mlx/backend/cuda/binary/subtract.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/binary/binary.cuh"
+
+namespace mlx::core {
+BINARY_GPU(Subtract)
+} // namespace mlx::core

mlx/backend/cuda/binary_two.cu

@@ -127,45 +127,99 @@ binary_two_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
+template <
+    typename Op,
+    typename In,
+    typename Out,
+    typename IdxT,
+    int NDIM,
+    int N_READS>
 __global__ void binary_two_g_nd(
     const In* a,
     const In* b,
     Out* out_a,
     Out* out_b,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
-  IdxT index = cg::this_grid().thread_rank();
+  auto block = cg::this_thread_block();
-  if (index < size) {
+  auto grid = cg::this_grid();
-    auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
+  IdxT index_rest =
-        index, shape.data(), a_strides.data(), b_strides.data());
+      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-    auto out = Op{}(a[a_idx], b[b_idx]);
+  if (index_rest >= size_rest) {
-    out_a[index] = out[0];
+    return;
-    out_b[index] = out[1];
   }
+
+  auto shape_x = shape[NDIM - 1];
+  auto a_stride_x = a_strides[NDIM - 1];
+  auto b_stride_x = b_strides[NDIM - 1];
+  IdxT index_x =
+      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
+      index_rest * shape_x, shape.data(), a_strides.data(), b_strides.data());
+  auto a_vec =
+      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, In(0));
+  auto b_vec =
+      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, In(0));
+
+  AlignedVector<Out, N_READS> out_vec_a;
+  AlignedVector<Out, N_READS> out_vec_b;
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    auto out = Op{}(a_vec[i], b_vec[i]);
+    out_vec_a[i] = out[0];
+    out_vec_b[i] = out[1];
+  }
+  store_vector(out_a + shape_x * index_rest, index_x, out_vec_a, shape_x);
+  store_vector(out_b + shape_x * index_rest, index_x, out_vec_b, shape_x);
 }
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void binary_two_g(
     const In* a,
     const In* b,
     Out* out_a,
     Out* out_b,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides a_strides,
     const __grid_constant__ Strides b_strides,
     int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
+  auto block = cg::this_thread_block();
-  if (index < size) {
+  auto grid = cg::this_grid();
-    auto [a_idx, b_idx] = elem_to_loc(
+  IdxT index_rest =
-        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
+      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-    auto out = Op{}(a[a_idx], b[b_idx]);
+  if (index_rest >= size_rest) {
-    out_a[index] = out[0];
+    return;
-    out_b[index] = out[1];
   }
+
+  auto shape_x = shape[ndim - 1];
+  auto a_stride_x = a_strides[ndim - 1];
+  auto b_stride_x = b_strides[ndim - 1];
+  IdxT index_x =
+      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  auto [a_idx, b_idx] = elem_to_loc(
+      index_rest * shape_x,
+      shape.data(),
+      a_strides.data(),
+      b_strides.data(),
+      ndim);
+  auto a_vec =
+      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, In(0));
+  auto b_vec =
+      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, In(0));
+
+  AlignedVector<Out, N_READS> out_vec_a;
+  AlignedVector<Out, N_READS> out_vec_b;
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    auto out = Op{}(a_vec[i], b_vec[i]);
+    out_vec_a[i] = out[0];
+    out_vec_b[i] = out[1];
+  }
+  store_vector(out_a + shape_x * index_rest, index_x, out_vec_a, shape_x);
+  store_vector(out_b + shape_x * index_rest, index_x, out_vec_b, shape_x);
 }
 
 template <typename Op, typename In, typename Out>
@@ -225,42 +279,64 @@ void binary_two_op_gpu_inplace(
                 auto& a_strides = strides[0];
                 auto& b_strides = strides[1];
                 int ndim = shape.size();
+                int work_per_thread = 1;
+                auto dim0 = ndim > 0 ? shape.back() : 1;
+                auto rest = out_a.size() / dim0;
+                if (dim0 >= 4) {
+                  work_per_thread = 4;
+                }
+                dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
+                auto block_dims = get_block_dims(dim0, rest, 1);
+                uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
+                uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
+
                 if (ndim <= 3) {
                   dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                    auto [num_blocks, block_dims] =
+                    auto kernel = cu::binary_two_g_nd<
-                        get_launch_args(out_a, large());
-                    encoder.add_kernel_node(
-                        cu::binary_two_g_nd<
                         Op,
                         InType,
                         OutType,
                         IdxT,
-                            dims_constant()>,
+                        dims_constant(),
-                        num_blocks,
+                        1>;
+                    if (work_per_thread == 4) {
+                      kernel = cu::binary_two_g_nd<
+                          Op,
+                          InType,
+                          OutType,
+                          IdxT,
+                          dims_constant(),
+                          4>;
+                    }
+                    encoder.add_kernel_node(
+                        kernel,
+                        {num_blocks_x, num_blocks_y},
                         block_dims,
                         0,
                         a.data<InType>(),
                         b.data<InType>(),
                         out_a.data<OutType>(),
                         out_b.data<OutType>(),
-                        out_a.size(),
+                        rest,
                         const_param<dims_constant()>(shape),
                         const_param<dims_constant()>(a_strides),
                         const_param<dims_constant()>(b_strides));
                   });
                 } else {
-                  auto [num_blocks, block_dims] =
+                  auto kernel = cu::binary_two_g<Op, InType, OutType, IdxT, 1>;
-                      get_launch_args(out_a, large());
+                  if (work_per_thread == 4) {
+                    kernel = cu::binary_two_g<Op, InType, OutType, IdxT, 4>;
+                  }
                   encoder.add_kernel_node(
-                      cu::binary_two_g<Op, InType, OutType, IdxT>,
+                      kernel,
-                      num_blocks,
+                      {num_blocks_x, num_blocks_y},
                       block_dims,
                       0,
                       a.data<InType>(),
                       b.data<InType>(),
                       out_a.data<OutType>(),
                       out_b.data<OutType>(),
-                      out_a.size(),
+                      rest,
                       const_param(shape),
                       const_param(a_strides),
                       const_param(b_strides),

mlx/backend/cuda/compiled.cpp

@@ -267,7 +267,8 @@ void Compiled::eval_gpu(
       }
     }
 
-    return std::make_pair(std::move(builder.os), std::move(kernel_names));
+    return std::make_tuple(
+        false, std::move(builder.os), std::move(kernel_names));
   });
 
   // Collapse contiguous dims to route to a faster kernel if possible. Also

mlx/backend/cuda/conv.cpp

@@ -1,18 +1,12 @@
 // Copyright © 2025 Apple Inc.
 
+#include "mlx/backend/cuda/conv/conv.h"
+#include "mlx/backend/cuda/cudnn_utils.h"
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/config.h"
 #include "mlx/backend/cuda/lru_cache.h"
 #include "mlx/backend/gpu/copy.h"
-#include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
 
-// cudnn_frontend.h redefines this macro.
-#undef CHECK_CUDA_ERROR
-
-#include <cudnn_frontend.h>
-#include <cudnn_frontend_find_plan.h>
-#include <fmt/format.h>
 #include <nvtx3/nvtx3.hpp>
 
 #include <cassert>
@@ -21,9 +15,6 @@ namespace mlx::core {
 
 namespace {
 
-// Not all engines support it so can not use this API now.
-#define MLX_USE_CUDNN_NATIVE_CUDA_GRAPH_API 0
-
 // Alias for better readability.
 #define CONV_FORWARD CUDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR
 #define CONV_BACKWARD_INPUT \
@@ -31,6 +22,9 @@ namespace {
 #define CONV_BACKWARD_WEIGHT \
   CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR
 
+// Custom placeholder representing fallback kernel.
+#define CONV_FALLBACK static_cast<cudnnBackendDescriptorType_t>(-1)
+
 struct ConvCacheKey {
   int device_id;
   cudnnDataType_t cudnn_dtype;
@@ -50,203 +44,13 @@ struct ConvCacheKey {
 auto& conv_cache() {
   static LRUBytesKeyCache<
       ConvCacheKey,
-      std::pair<cudnnBackendDescriptorType_t, cudnn_frontend::ExecutionPlan>>
+      std::pair<
-      cache(/* capacity */ 128);
+          cudnnBackendDescriptorType_t,
+          std::optional<cudnn_frontend::ExecutionPlan>>>
+      cache("MLX_CUDA_CONV_CACHE_SIZE", /* default_capacity */ 128);
   return cache;
 }
 
-template <typename T, typename Vec>
-inline SmallVector<T> convert_vector(const Vec& vec) {
-  return SmallVector<T>(vec.begin(), vec.end());
-}
-
-template <typename T, template <typename U> class Vec>
-inline std::array<T, MAX_NDIM> fixed_vector(const Vec<T>& vec) {
-  if (vec.size() > MAX_NDIM) {
-    throw std::runtime_error(
-        fmt::format("ndim can not be larger than {}.", MAX_NDIM));
-  }
-  std::array<T, MAX_NDIM> result = {};
-  std::copy_n(vec.begin(), vec.size(), result.begin());
-  return result;
-}
-
-auto nhwc_to_nchw(const array& x) {
-  auto shape = convert_vector<int64_t>(x.shape());
-  shape.insert(shape.begin() + 1, shape.back());
-  shape.erase(shape.end() - 1);
-  auto strides = convert_vector<int64_t>(x.strides());
-  strides.insert(strides.begin() + 1, strides.back());
-  strides.erase(strides.end() - 1);
-  return std::make_tuple(std::move(shape), std::move(strides));
-}
-
-inline cudnnDataType_t dtype_to_cudnn_type(Dtype dtype) {
-  switch (dtype) {
-    case int8:
-      return CUDNN_DATA_INT8;
-    case int32:
-      return CUDNN_DATA_INT32;
-    case uint8:
-      return CUDNN_DATA_UINT8;
-    case float16:
-      return CUDNN_DATA_HALF;
-    case bfloat16:
-      return CUDNN_DATA_BFLOAT16;
-    case float32:
-      return CUDNN_DATA_FLOAT;
-    case float64:
-      return CUDNN_DATA_DOUBLE;
-    default:
-      throw std::runtime_error(fmt::format(
-          "Unsupported dtype in Convolution: {}.", dtype_to_string(dtype)));
-  }
-}
-
-inline uint8_t get_alignment(const array& x) {
-  uint8_t alignment = 1;
-  uintptr_t address = reinterpret_cast<uintptr_t>(x.data<void>());
-  for (; alignment < 32; alignment *= 2) {
-    if (address % (alignment * 2)) {
-      return alignment;
-    }
-  }
-  return alignment;
-}
-
-inline cudnn_frontend::Tensor build_tensor(int64_t id, const array& x) {
-  auto [shape, strides] = nhwc_to_nchw(x);
-  return cudnn_frontend::TensorBuilder()
-      .setDim(shape.size(), shape.data())
-      .setStrides(strides.size(), strides.data())
-      .setId(id)
-      .setAlignment(get_alignment(x))
-      .setDataType(dtype_to_cudnn_type(x.dtype()))
-      .build();
-}
-
-cudnn_frontend::EngineConfigList get_engine_configs(
-    cudnnBackendDescriptorType_t backend_type,
-    Dtype dtype,
-    cudnn_frontend::OperationGraph& op_graph,
-    bool use_fallback = false) {
-  cudnn_frontend::GeneratorSource source;
-  if (use_fallback) {
-    source = [&backend_type](cudnn_frontend::OperationGraph& op_graph) {
-      auto fallback = cudnn_frontend::EngineFallbackListBuilder()
-                          .setOperationGraph(op_graph)
-                          .setOperation(backend_type)
-                          .build();
-      return fallback.getFallbackList();
-    };
-  } else {
-    source = [](cudnn_frontend::OperationGraph& op_graph) {
-      auto heuristics = cudnn_frontend::EngineHeuristicsBuilder()
-                            .setOperationGraph(op_graph)
-                            .setHeurMode(CUDNN_HEUR_MODE_A)
-                            .build();
-      return heuristics.getEngineConfig(heuristics.getEngineConfigCount());
-    };
-  }
-
-  cudnn_frontend::EngineConfigGenerator generator(1, &source);
-  auto configs = generator.generate_engine_config(op_graph);
-
-  cudnn_frontend::EngineConfigList filtered_configs;
-  cudnn_frontend::filter(configs, filtered_configs, [dtype](auto c) {
-    if (cudnn_frontend::hasNumericalNote<
-            CUDNN_NUMERICAL_NOTE_DOWN_CONVERT_INPUTS>(c)) {
-      return true;
-    }
-    if (cudnn_frontend::hasNumericalNote<CUDNN_NUMERICAL_NOTE_TENSOR_CORE>(c) &&
-        dtype == float32 && !env::enable_tf32()) {
-      return true;
-    }
-    return false;
-  });
-  return filtered_configs;
-}
-
-bool execute_plan(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    array& x,
-    array& w,
-    array& y) {
-  int workspace_size = plan.getWorkspaceSize();
-  array workspace(allocator::malloc(workspace_size), {workspace_size}, uint8);
-
-  int64_t uids[3] = {'x', 'w', 'y'};
-  void* data_ptrs[3] = {
-      x.data<void>(),
-      w.data<void>(),
-      y.data<void>(),
-  };
-
-  auto variantPack = cudnn_frontend::VariantPackBuilder()
-                         .setWorkspacePointer(workspace.data<void>())
-                         .setDataPointers(3, data_ptrs)
-                         .setUids(3, uids)
-                         .build();
-
-  auto handle = encoder.device().cudnn_handle();
-  cudnnSetStream(handle, encoder.stream());
-
-#if CUDNN_VERSION >= 90500 && MLX_USE_CUDNN_NATIVE_CUDA_GRAPH_API
-  cudaGraph_t graph;
-  cudaGraphCreate(&graph, 0);
-  std::unique_ptr<cudaGraph_t, void (*)(cudaGraph_t*)> graph_freer(
-      &graph, [](cudaGraph_t* p) { cudaGraphDestroy(*p); });
-  if (cudnnBackendPopulateCudaGraph(
-          handle, plan.get_raw_desc(), variantPack.get_raw_desc(), graph) !=
-      CUDNN_STATUS_SUCCESS) {
-    return false;
-  }
-  encoder.add_graph_node(graph);
-#else
-  auto capture = encoder.capture_context();
-  if (cudnnBackendExecute(
-          handle, plan.get_raw_desc(), variantPack.get_raw_desc()) !=
-      CUDNN_STATUS_SUCCESS) {
-    // Discard the captured graph when failed.
-    capture.discard = true;
-    return false;
-  }
-#endif
-
-  encoder.add_temporary(workspace);
-  return true;
-}
-
-bool try_engines(
-    cu::CommandEncoder& encoder,
-    const ConvCacheKey& cache_key,
-    cudnnBackendDescriptorType_t backend_type,
-    cudnn_frontend::EngineConfigList& configs,
-    const std::string& op_graph_tag,
-    array& x,
-    array& w,
-    array& y) {
-  for (auto& config : configs) {
-    try {
-      auto plan = cudnn_frontend::ExecutionPlanBuilder()
-                      .setHandle(encoder.device().cudnn_handle())
-                      .setEngineConfig(config, op_graph_tag)
-                      .build();
-      if (execute_plan(encoder, plan, x, w, y)) {
-        conv_cache().emplace(
-            cache_key, std::make_pair(backend_type, std::move(plan)));
-        return true;
-      }
-    } catch (cudnn_frontend::cudnnException& error) {
-      if (error.getCudnnStatus() != CUDNN_STATUS_NOT_SUPPORTED) {
-        throw;
-      }
-    }
-  }
-  return false;
-}
-
 auto get_conv_op_settings(
     cudnnBackendDescriptorType_t backend_type,
     array& x,
@@ -291,7 +95,7 @@ auto get_conv_op_settings(
   }
 }
 
-std::optional<cudnn_frontend::OperationGraph> build_op_graph(
+std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
     cu::CommandEncoder& encoder,
     cudnnBackendDescriptorType_t backend_type,
     Dtype dtype,
@@ -317,9 +121,9 @@ std::optional<cudnn_frontend::OperationGraph> build_op_graph(
                          .build();
 
     auto op = cudnn_frontend::OperationBuilder(backend_type)
-                  .setxDesc(build_tensor('x', x))
+                  .setxDesc(build_cudnn_tensor_nchw('x', x))
-                  .setwDesc(build_tensor('w', w))
+                  .setwDesc(build_cudnn_tensor_nchw('w', w))
-                  .setyDesc(build_tensor('y', y))
+                  .setyDesc(build_cudnn_tensor_nchw('y', y))
                   .setcDesc(conv_desc)
                   .build();
 
@@ -336,6 +140,42 @@ std::optional<cudnn_frontend::OperationGraph> build_op_graph(
   }
 }
 
+// Transpose from (C_out, H, W, C_in / groups) to (C_in, H, W, C_out / groups).
+array group_transpose(
+    const array& x,
+    int groups,
+    int group_dim,
+    int axis1,
+    int axis2,
+    Stream s) {
+  if (groups == 1) {
+    return swapaxes_in_eval(x, axis1, axis2);
+  }
+  int ndim = x.ndim();
+  if (group_dim < 0) {
+    group_dim += ndim;
+  }
+  if (axis1 < 0) {
+    axis1 += ndim;
+  }
+  if (axis2 < 0) {
+    axis2 += ndim;
+  }
+  if (group_dim <= axis1) {
+    axis1 += 1;
+  }
+  if (group_dim <= axis2) {
+    axis2 += 1;
+  }
+  auto shape = x.shape();
+  shape.insert(shape.begin() + group_dim, groups);
+  shape[group_dim + 1] = shape[group_dim + 1] / groups;
+  array x_trans = reshape_in_eval(x, std::move(shape), s);
+  x_trans = swapaxes_in_eval(x_trans, axis1, axis2);
+  x_trans = flatten_in_eval(x_trans, group_dim, group_dim + 1, s);
+  return x_trans;
+}
+
 // Do necessary transposes and copies to prepare the inputs and outputs for
 // building the cuDNN conv op. It is safe to be called multiple times in one
 // eval_gpu, with cost of possible redundant copies.
@@ -345,13 +185,14 @@ std::tuple<array, array, array> prepare_args(
     array in,
     array wt,
     array out,
+    int groups,
     Stream s) {
   // Transpose the args depending on the backend type.
   // TODO: Handle groups.
   if (backend_type == CONV_BACKWARD_INPUT) {
-    wt = swapaxes_in_eval(wt, 0, -1);
+    wt = group_transpose(wt, groups, 0, 0, -1, s);
   } else if (backend_type == CONV_BACKWARD_WEIGHT) {
-    in = swapaxes_in_eval(in, 0, -1);
+    in = group_transpose(in, groups, -1, 0, -1, s);
     wt = swapaxes_in_eval(wt, 0, -1);
     // Create a contiguous array that shares the data with |out|, but with dim
     // C_in and C_out swapped.
@@ -444,12 +285,12 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   ConvCacheKey cache_key{
       encoder.device().cuda_device(),
       dtype_to_cudnn_type(dtype),
-      fixed_vector(in.shape()),
+      vector_key(in.shape()),
-      fixed_vector(wt.shape()),
+      vector_key(wt.shape()),
-      fixed_vector(kernel_strides_),
+      vector_key(kernel_strides_),
-      fixed_vector(padding_lo_),
+      vector_key(padding_lo_),
-      fixed_vector(padding_hi_),
+      vector_key(padding_hi_),
-      fixed_vector(kernel_dilation_),
+      vector_key(kernel_dilation_),
       groups_,
       flip_,
       get_alignment(in),
@@ -457,12 +298,30 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
       get_alignment(out)};
   if (auto it = conv_cache().find(cache_key); it != conv_cache().end()) {
     auto& [backend_type, plan] = it->second;
-    std::tie(in, wt, out) = prepare_args(encoder, backend_type, in, wt, out, s);
+    if (plan) {
+      // Run cached plan.
+      std::tie(in, wt, out) =
+          prepare_args(encoder, backend_type, in, wt, out, groups_, s);
       register_args(encoder, backend_type, in, wt, out, out_);
       auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-    if (!execute_plan(encoder, plan, x, w, y)) {
+      if (!encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
         throw std::runtime_error("[conv] Cached plan failed to execute.");
       }
+    } else {
+      // Run fallback kernel.
+      gemm_conv(
+          encoder,
+          in,
+          wt,
+          out,
+          kernel_strides_,
+          padding_lo_,
+          kernel_dilation_,
+          input_dilation_,
+          groups_,
+          flip_,
+          s);
+    }
     return;
   }
 
@@ -490,7 +349,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   std::optional<cudnn_frontend::OperationGraph> op_graph;
   for (auto try_backend : try_backends) {
     auto [in_copy, wt_copy, out_copy] =
-        prepare_args(encoder, try_backend, in, wt, out, s);
+        prepare_args(encoder, try_backend, in, wt, out, groups_, s);
     auto [x, w, y] = dispatch_args(try_backend, in_copy, wt_copy, out_copy);
     auto [stride, padding_lo, padding_hi, dilation] = get_conv_op_settings(
         try_backend,
@@ -502,7 +361,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
         padding_hi_,
         kernel_dilation_,
         input_dilation_);
-    op_graph = build_op_graph(
+    op_graph = build_conv_op_graph(
         encoder,
         try_backend,
         dtype,
@@ -521,26 +380,39 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
       break;
     }
   }
-  if (!op_graph) {
-    throw std::runtime_error("[conv] Can not build op graph.");
-  }
 
-  // Get ready to execute the graph.
+  if (op_graph) {
+    // Setup inputs and outputs.
     register_args(encoder, backend_type, in, wt, out, out_);
 
-  // Try to run plans based on heuristics.
+    // Find a plan for the graph and execute it.
-  auto configs = get_engine_configs(backend_type, dtype, *op_graph);
+    auto plan = find_cudnn_plan_from_op_graph(
-  auto tag = op_graph->getTag();
+        encoder.device().cudnn_handle(), backend_type, dtype, *op_graph);
+    if (!plan) {
+      throw std::runtime_error("[conv] Unable to find an execution plan.");
+    }
     auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
+    if (encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
+      conv_cache().emplace(
+          cache_key, std::make_pair(backend_type, std::move(*plan)));
       return;
     }
-  // Then try fallback plans.
-  configs = get_engine_configs(backend_type, dtype, *op_graph);
-  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
-    return;
   }
-  throw std::runtime_error("[conv] Unable to find a working engine.");
+
+  // Use fallback kernel for settings not supported by cuDNN.
+  gemm_conv(
+      encoder,
+      in,
+      wt,
+      out,
+      kernel_strides_,
+      padding_lo_,
+      kernel_dilation_,
+      input_dilation_,
+      groups_,
+      flip_,
+      s);
+  conv_cache().emplace(cache_key, std::make_pair(CONV_FALLBACK, std::nullopt));
 }
 
 } // namespace mlx::core

mlx/backend/cuda/conv/conv.h

@@ -0,0 +1,126 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/gpu/copy.h"
+
+namespace mlx::core {
+
+template <int NDIM>
+struct ConvParams {
+  int N; // Batch size
+  int C; // In channels
+  int O; // Out channels
+  int strides[NDIM];
+  int padding[NDIM];
+  int kernel_dilation[NDIM];
+  int input_dilation[NDIM];
+  int groups;
+  bool flip;
+  int in_spatial_dims[NDIM];
+  int wt_spatial_dims[NDIM];
+  int out_spatial_dims[NDIM];
+  int64_t in_strides[NDIM + 2];
+
+  ConvParams(
+      const array& in,
+      const array& wt,
+      const array& out,
+      const std::vector<int>& strides,
+      const std::vector<int>& padding,
+      const std::vector<int>& kernel_dilation,
+      const std::vector<int>& input_dilation,
+      int groups,
+      bool flip)
+      : N(in.shape(0)),
+        C(in.shape(-1)),
+        O(wt.shape(0)),
+        groups(groups),
+        flip(flip) {
+    std::copy_n(strides.begin(), NDIM, this->strides);
+    std::copy_n(padding.begin(), NDIM, this->padding);
+    std::copy_n(kernel_dilation.begin(), NDIM, this->kernel_dilation);
+    std::copy_n(input_dilation.begin(), NDIM, this->input_dilation);
+    std::copy_n(in.shape().begin() + 1, NDIM, this->in_spatial_dims);
+    std::copy_n(wt.shape().begin() + 1, NDIM, this->wt_spatial_dims);
+    std::copy_n(out.shape().begin() + 1, NDIM, this->out_spatial_dims);
+    std::copy_n(in.strides().begin(), NDIM + 2, this->in_strides);
+  }
+};
+
+void gemm_grouped_conv(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    const array& wt,
+    array& out,
+    const std::vector<int>& strides,
+    const std::vector<int>& padding,
+    const std::vector<int>& kernel_dilation,
+    const std::vector<int>& input_dilation,
+    int groups,
+    bool flip,
+    Stream s);
+
+void gemm_conv(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    const array& wt,
+    array& out,
+    const std::vector<int>& strides,
+    const std::vector<int>& padding,
+    const std::vector<int>& kernel_dilation,
+    const std::vector<int>& input_dilation,
+    bool flip,
+    Stream s);
+
+inline void gemm_conv(
+    cu::CommandEncoder& encoder,
+    array in,
+    array wt,
+    array& out,
+    const std::vector<int>& strides,
+    const std::vector<int>& padding,
+    const std::vector<int>& kernel_dilation,
+    const std::vector<int>& input_dilation,
+    int groups,
+    bool flip,
+    Stream s) {
+  if (!in.flags().row_contiguous) {
+    in = contiguous_copy_gpu(in, s);
+    encoder.add_temporary(in);
+  }
+  if (!wt.flags().row_contiguous) {
+    wt = contiguous_copy_gpu(wt, s);
+    encoder.add_temporary(wt);
+  }
+
+  if (groups == 1) {
+    gemm_conv(
+        encoder,
+        in,
+        wt,
+        out,
+        strides,
+        padding,
+        kernel_dilation,
+        input_dilation,
+        flip,
+        s);
+  } else {
+    gemm_grouped_conv(
+        encoder,
+        in,
+        wt,
+        out,
+        strides,
+        padding,
+        kernel_dilation,
+        input_dilation,
+        groups,
+        flip,
+        s);
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cuda/conv/gemm_conv.cu

@@ -0,0 +1,217 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/conv/conv.h"
+#include "mlx/backend/cuda/gemms/cublas_gemm.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename T, int NDIM>
+__global__ void naive_unfold_nd(
+    const T* in,
+    T* out,
+    int filter_size,
+    int out_pixels,
+    const __grid_constant__ ConvParams<NDIM> params) {
+  auto block = cg::this_thread_block();
+  auto tid = block.group_index();
+  auto lid = block.thread_index();
+
+  int index_batch = tid.z / out_pixels; // [0, N)
+  int index_out_spatial = tid.z % out_pixels; // [0, H_out * W_out)
+  int index_wt_spatial =
+      tid.x * block.dim_threads().x + lid.x; // [0, H_wt * W_wt)
+
+  if (index_wt_spatial >= filter_size / params.C) {
+    return;
+  }
+
+  in += tid.y; // [0, C)
+  out += tid.z * filter_size + index_wt_spatial * params.C + tid.y;
+
+  bool valid = index_batch < params.N;
+
+  // Get the coordinates in input.
+  int index_in[NDIM] = {};
+#pragma unroll
+  for (int i = NDIM - 1; i >= 0; --i) {
+    int index_out = index_out_spatial % params.out_spatial_dims[i];
+    int index_wt = index_wt_spatial % params.wt_spatial_dims[i];
+
+    if (params.flip) {
+      index_wt = params.wt_spatial_dims[i] - index_wt - 1;
+    }
+
+    int index = index_out * params.strides[i] - params.padding[i] +
+        index_wt * params.kernel_dilation[i];
+    int index_max =
+        1 + params.input_dilation[i] * (params.in_spatial_dims[i] - 1);
+
+    valid &= (index >= 0) && (index < index_max) &&
+        (index % params.input_dilation[i] == 0);
+
+    index_in[i] = index / params.input_dilation[i];
+
+    index_out_spatial /= params.out_spatial_dims[i];
+    index_wt_spatial /= params.wt_spatial_dims[i];
+  }
+
+  if (valid) {
+    int in_offset = index_batch * params.in_strides[0];
+#pragma unroll
+    for (int i = 0; i < NDIM; ++i) {
+      in_offset += index_in[i] * params.in_strides[i + 1];
+    }
+    *out = in[in_offset];
+  } else {
+    *out = T{0};
+  }
+}
+
+} // namespace cu
+
+template <int NDIM>
+array unfold_inputs_nd(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    int mat_M,
+    int mat_K,
+    int mat_N,
+    ConvParams<NDIM>& params) {
+  array unfolded({mat_M, mat_K}, in.dtype(), nullptr, {});
+  unfolded.set_data(allocator::malloc(unfolded.nbytes()));
+  encoder.add_temporary(unfolded);
+
+  int filter_size = params.C;
+#pragma unroll
+  for (int i = 0; i < NDIM; ++i) {
+    filter_size *= params.wt_spatial_dims[i];
+  }
+
+  int out_pixels = 1;
+#pragma unroll
+  for (int i = 0; i < NDIM; ++i) {
+    out_pixels *= params.out_spatial_dims[i];
+  }
+
+  int wt_spatial_size = mat_K / params.C;
+  dim3 block_dims;
+  block_dims.x = std::min(std::max(wt_spatial_size, 32), 1024);
+  dim3 num_blocks;
+  num_blocks.x = cuda::ceil_div(wt_spatial_size, block_dims.x);
+  num_blocks.y = params.C;
+  num_blocks.z = mat_M;
+
+  encoder.set_input_array(in);
+  encoder.set_output_array(unfolded);
+  dispatch_float_types(in.dtype(), "unfold", [&](auto type_tag) {
+    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    encoder.add_kernel_node(
+        cu::naive_unfold_nd<DataType, NDIM>,
+        num_blocks,
+        block_dims,
+        0,
+        in.data<DataType>(),
+        unfolded.data<DataType>(),
+        filter_size,
+        out_pixels,
+        params);
+  });
+
+  return unfolded;
+}
+
+template <int NDIM>
+void gemm_conv_nd(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    const array& wt,
+    array& out,
+    ConvParams<NDIM>& params,
+    Stream s) {
+  // Get gemm shapes.
+  int mat_M = out.size() / params.O; // N * H_out * W_out
+  int mat_K = wt.size() / params.O; // C * H_wt * W_wt
+  int mat_N = params.O; // O
+
+  // Unfold input to (N * H_out * W_out, C * H_wt * W_wt) for gemm.
+  array in_unfolded =
+      unfold_inputs_nd<NDIM>(encoder, in, mat_M, mat_K, mat_N, params);
+
+  // Reshape weight to (C * H_wt * W_wt, O) for gemm.
+  array wt_reshaped({mat_K, mat_N}, wt.dtype(), nullptr, {});
+  wt_reshaped.copy_shared_buffer(
+      wt,
+      {1, mat_K},
+      {false, false, /* col_contiguous */ true},
+      wt.data_size());
+
+  // Single batch.
+  Shape batch_shape{1};
+  Strides a_batch_strides{0};
+  Strides b_batch_strides{0};
+
+  // Run matmul.
+  CublasGemm gemm(
+      encoder.device(),
+      in.dtype(),
+      false, // a_transposed
+      mat_M, // a_rows
+      mat_K, // a_cols
+      mat_K, // lda
+      true, // b_transposed
+      mat_K, // b_rows
+      mat_N, // b_cols
+      mat_K, // ldb
+      batch_shape.back(),
+      a_batch_strides.back(),
+      b_batch_strides.back());
+  gemm.run(
+      encoder,
+      out,
+      in_unfolded,
+      wt_reshaped,
+      batch_shape,
+      a_batch_strides,
+      b_batch_strides);
+}
+
+void gemm_conv(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    const array& wt,
+    array& out,
+    const std::vector<int>& strides,
+    const std::vector<int>& padding,
+    const std::vector<int>& kernel_dilation,
+    const std::vector<int>& input_dilation,
+    bool flip,
+    Stream s) {
+  int conv_ndim = in.ndim() - 2;
+  if (conv_ndim < 1 || conv_ndim > 3) {
+    throw std::runtime_error(
+        fmt::format("[conv] Unsupported gemm_conv for {}D conv.", conv_ndim));
+  }
+  dispatch_1_2_3(conv_ndim, [&](auto ndim_constant) {
+    ConvParams<ndim_constant()> params(
+        in,
+        wt,
+        out,
+        strides,
+        padding,
+        kernel_dilation,
+        input_dilation,
+        1, // groups
+        flip);
+    gemm_conv_nd<ndim_constant()>(encoder, in, wt, out, params, s);
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/conv/gemm_grouped_conv.cu

@@ -0,0 +1,231 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/conv/conv.h"
+#include "mlx/backend/cuda/gemms/cublas_gemm.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename T, int NDIM>
+__global__ void naive_grouped_unfold_transpose_nd(
+    const T* in,
+    T* out,
+    int filter_size,
+    int out_pixels,
+    const __grid_constant__ ConvParams<NDIM> params) {
+  auto block = cg::this_thread_block();
+  auto tid = block.group_index();
+  auto lid = block.thread_index();
+
+  int index_batch = tid.z / out_pixels; // [0, N)
+  int index_out_spatial = tid.z % out_pixels; // [0, H_out * W_out)
+  int index_wt_spatial =
+      tid.x * block.dim_threads().x + lid.x; // [0, H_wt * W_wt)
+
+  if (index_wt_spatial >= filter_size / params.C) {
+    return;
+  }
+
+  in += tid.y; // [0, C)
+  out += tid.z * filter_size + tid.y * (filter_size / params.C);
+
+  bool valid = index_batch < params.N;
+
+  // Get the coordinates in input.
+  int index_in[NDIM] = {};
+  int wt_stride = 1;
+#pragma unroll
+  for (int i = NDIM - 1; i >= 0; --i) {
+    int index_out = index_out_spatial % params.out_spatial_dims[i];
+    int index_wt = index_wt_spatial % params.wt_spatial_dims[i];
+    out += index_wt * wt_stride;
+
+    if (params.flip) {
+      index_wt = params.wt_spatial_dims[i] - index_wt - 1;
+    }
+
+    int index = index_out * params.strides[i] - params.padding[i] +
+        index_wt * params.kernel_dilation[i];
+    int index_max =
+        1 + params.input_dilation[i] * (params.in_spatial_dims[i] - 1);
+
+    valid &= (index >= 0) && (index < index_max) &&
+        (index % params.input_dilation[i] == 0);
+
+    index_in[i] = index / params.input_dilation[i];
+
+    index_out_spatial /= params.out_spatial_dims[i];
+    index_wt_spatial /= params.wt_spatial_dims[i];
+    wt_stride *= params.wt_spatial_dims[i];
+  }
+
+  if (valid) {
+    int in_offset = index_batch * params.in_strides[0];
+#pragma unroll
+    for (int i = 0; i < NDIM; ++i) {
+      in_offset += index_in[i] * params.in_strides[i + 1];
+    }
+    *out = in[in_offset];
+  } else {
+    *out = T{0};
+  }
+}
+
+} // namespace cu
+
+template <int NDIM>
+array grouped_unfold_transpose_inputs_nd(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    int mat_M,
+    int mat_K,
+    int mat_N,
+    ConvParams<NDIM>& params) {
+  array unfolded({mat_M, mat_K * params.groups}, in.dtype(), nullptr, {});
+  unfolded.set_data(allocator::malloc(unfolded.nbytes()));
+  encoder.add_temporary(unfolded);
+
+  int filter_size = params.C;
+#pragma unroll
+  for (int i = 0; i < NDIM; ++i) {
+    filter_size *= params.wt_spatial_dims[i];
+  }
+
+  int out_pixels = 1;
+#pragma unroll
+  for (int i = 0; i < NDIM; ++i) {
+    out_pixels *= params.out_spatial_dims[i];
+  }
+
+  int wt_spatial_size = (mat_K * params.groups) / params.C;
+  dim3 block_dims;
+  block_dims.x = std::min(std::max(wt_spatial_size, 32), 1024);
+  dim3 num_blocks;
+  num_blocks.x = cuda::ceil_div(wt_spatial_size, block_dims.x);
+  num_blocks.y = params.C;
+  num_blocks.z = mat_M;
+
+  encoder.set_input_array(in);
+  encoder.set_output_array(unfolded);
+  dispatch_float_types(in.dtype(), "unfold", [&](auto type_tag) {
+    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    encoder.add_kernel_node(
+        cu::naive_grouped_unfold_transpose_nd<DataType, NDIM>,
+        num_blocks,
+        block_dims,
+        0,
+        in.data<DataType>(),
+        unfolded.data<DataType>(),
+        filter_size,
+        out_pixels,
+        params);
+  });
+
+  return unfolded;
+}
+
+template <int NDIM>
+void gemm_grouped_conv_nd(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    const array& wt,
+    array& out,
+    ConvParams<NDIM>& params,
+    Stream s) {
+  // Get gemm shapes.
+  int C_per_group = params.C / params.groups;
+  int O_per_group = params.O / params.groups;
+  int mat_M = out.size() / params.O; // N * H_out * W_out
+  int mat_K = wt.size() / params.O; // C_per_group * H_wt * W_wt
+  int mat_N = O_per_group; // O_per_group
+
+  // Unfold input to (N * H_out * W_out, C * H_wt * W_wt) for gemm.
+  array in_unfolded = grouped_unfold_transpose_inputs_nd<NDIM>(
+      encoder, in, mat_M, mat_K, mat_N, params);
+
+  // Reshape weight to (O, C_per_group, H_wt * W_wt) for gemm.
+  int wt_spatial_size = (wt.size() / wt.shape(0)) / wt.shape(-1);
+  array wt_view(
+      {params.O, C_per_group, wt_spatial_size}, wt.dtype(), nullptr, {});
+  wt_view.copy_shared_buffer(
+      wt, {wt.strides(0), 1, C_per_group}, wt.flags(), wt.size());
+  array wt_reshaped = contiguous_copy_gpu(wt_view, s);
+
+  // Batch with size of groups.
+  Shape batch_shape{params.groups};
+  Strides a_batch_strides{mat_K};
+  Strides b_batch_strides{mat_N * mat_K};
+
+  // Run matmul.
+  CublasGemm gemm(
+      encoder.device(),
+      in.dtype(),
+      false, // a_transposed
+      mat_M, // a_rows
+      mat_K, // a_cols
+      mat_K * params.groups, // lda
+      true, // b_transposed
+      mat_K, // b_rows
+      mat_N, // b_cols
+      mat_K, // ldb
+      batch_shape.back(),
+      a_batch_strides.back(),
+      b_batch_strides.back());
+  gemm.set_out(
+      out.dtype(),
+      false, // out_transposed
+      mat_M, // out_rows
+      mat_N, // out_cols
+      mat_N * params.groups, // out_ld
+      params.groups, // batch_count
+      mat_N); // batch_stride
+  gemm.run(
+      encoder,
+      out,
+      in_unfolded,
+      wt_reshaped,
+      batch_shape,
+      a_batch_strides,
+      b_batch_strides);
+}
+
+void gemm_grouped_conv(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    const array& wt,
+    array& out,
+    const std::vector<int>& strides,
+    const std::vector<int>& padding,
+    const std::vector<int>& kernel_dilation,
+    const std::vector<int>& input_dilation,
+    int groups,
+    bool flip,
+    Stream s) {
+  int conv_ndim = in.ndim() - 2;
+  if (conv_ndim < 1 || conv_ndim > 3) {
+    throw std::runtime_error(
+        fmt::format("[conv] Unsupported gemm_conv for {}D conv.", conv_ndim));
+  }
+  dispatch_1_2_3(conv_ndim, [&](auto ndim_constant) {
+    ConvParams<ndim_constant()> params(
+        in,
+        wt,
+        out,
+        strides,
+        padding,
+        kernel_dilation,
+        input_dilation,
+        groups,
+        flip);
+    gemm_grouped_conv_nd<ndim_constant()>(encoder, in, wt, out, params, s);
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/copy.cu

@@ -15,8 +15,8 @@ void copy_gpu_inplace(
     int64_t offset_out,
     CopyType ctype,
     const Stream& s,
-    const std::optional<array>& dynamic_offset_in,
+    std::optional<array> dynamic_offset_in,
-    const std::optional<array>& dynamic_offset_out) {
+    std::optional<array> dynamic_offset_out) {
   if (out.size() == 0) {
     return;
   }
@@ -44,6 +44,16 @@ void copy_gpu_inplace(
           strides_vec[0]);
     } else {
       if (dynamic_offset_in || dynamic_offset_out) {
+        if (!dynamic_offset_in) {
+          dynamic_offset_in = array(0, int64);
+          encoder.add_temporary(*dynamic_offset_in);
+        }
+        if (!dynamic_offset_out) {
+          dynamic_offset_out = array(0, int64);
+          encoder.add_temporary(*dynamic_offset_out);
+        }
+        encoder.set_input_array(*dynamic_offset_in);
+        encoder.set_input_array(*dynamic_offset_out);
         copy_general_dynamic(
             encoder,
             ctype,
@@ -54,8 +64,8 @@ void copy_gpu_inplace(
             shape_collapsed,
             strides_vec[0],
             strides_vec[1],
-            dynamic_offset_in ? *dynamic_offset_in : array(0, int64),
+            *dynamic_offset_in,
-            dynamic_offset_out ? *dynamic_offset_out : array(0, int64));
+            *dynamic_offset_out);
       } else {
         copy_general(
             encoder,

mlx/backend/cuda/copy/copy_general.cu

@@ -10,37 +10,80 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT, int NDIM>
+template <typename In, typename Out, typename IdxT, int NDIM, int N_READS>
 __global__ void copy_gg_nd(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_in,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_out) {
-  IdxT index = cg::this_grid().thread_rank();
+  auto block = cg::this_thread_block();
-  if (index < size) {
+  auto grid = cg::this_grid();
-    auto [idx_in, idx_out] = elem_to_loc_nd<NDIM>(
+  IdxT index_rest =
-        index, shape.data(), strides_in.data(), strides_out.data());
+      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
+  if (index_rest >= size_rest) {
+    return;
   }
+
+  auto shape_x = shape[NDIM - 1];
+  auto in_stride_x = strides_in[NDIM - 1];
+  auto out_stride_x = strides_out[NDIM - 1];
+  IdxT index_x =
+      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  auto [idx_in, idx_out] = elem_to_loc_nd<NDIM>(
+      index_rest * shape_x,
+      shape.data(),
+      strides_in.data(),
+      strides_out.data());
+
+  auto in_vec =
+      load_vector<N_READS>(in + idx_in, index_x, shape_x, in_stride_x, In(0));
+  AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    out_vec[i] = CastOp<In, Out>{}(in_vec[i]);
+  }
+  store_vector(out + idx_out, index_x, out_vec, shape_x, out_stride_x);
 }
 
-template <typename In, typename Out, typename IdxT>
+template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_gg(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides strides_in,
     const __grid_constant__ Strides strides_out,
     int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
+  auto block = cg::this_thread_block();
-  if (index < size) {
+  auto grid = cg::this_grid();
-    auto [idx_in, idx_out] = elem_to_loc(
+  IdxT index_rest =
-        index, shape.data(), strides_in.data(), strides_out.data(), ndim);
+      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
+  if (index_rest >= size_rest) {
+    return;
   }
+
+  auto shape_x = shape[ndim - 1];
+  auto in_stride_x = strides_in[ndim - 1];
+  auto out_stride_x = strides_out[ndim - 1];
+  IdxT index_x =
+      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  auto [idx_in, idx_out] = elem_to_loc(
+      index_rest * shape_x,
+      shape.data(),
+      strides_in.data(),
+      strides_out.data(),
+      ndim);
+
+  auto in_vec =
+      load_vector<N_READS>(in + idx_in, index_x, shape_x, in_stride_x, In(0));
+  AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    out_vec[i] = CastOp<In, Out>{}(in_vec[i]);
+  }
+  store_vector(out + idx_out, index_x, out_vec, shape_x, out_stride_x);
 }
 
 } // namespace cu
@@ -69,33 +112,52 @@ void copy_general(
             size_t data_size = 1;
             for (auto& s : shape)
               data_size *= s;
+
+            int work_per_thread = 1;
+            auto dim0 = ndim > 0 ? shape.back() : 1;
+            auto rest = data_size / dim0;
+            if (dim0 >= 4) {
+              work_per_thread = 4;
+            }
+
+            dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
+            auto block_dims = get_block_dims(dim0, rest, 1);
+            uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
+            uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
+
             if (ndim <= 3) {
               dispatch_1_2_3(ndim, [&](auto ndim_constant) {
-                auto [num_blocks, block_dims] =
+                auto kernel =
-                    get_launch_args(data_size, shape, out.strides(), large());
+                    cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant(), 1>;
+                if (work_per_thread == 4) {
+                  kernel =
+                      cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant(), 4>;
+                }
                 encoder.add_kernel_node(
-                    cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant()>,
+                    kernel,
-                    num_blocks,
+                    {num_blocks_x, num_blocks_y},
                     block_dims,
                     0,
                     in_ptr,
                     out_ptr,
-                    data_size,
+                    rest,
                     const_param<ndim_constant()>(shape),
                     const_param<ndim_constant()>(strides_in),
                     const_param<ndim_constant()>(strides_out));
               });
             } else { // ndim >= 4
-              auto [num_blocks, block_dims] =
+              auto kernel = cu::copy_gg<InType, OutType, IdxT, 1>;
-                  get_launch_args(data_size, shape, out.strides(), large());
+              if (work_per_thread == 4) {
+                kernel = cu::copy_gg<InType, OutType, IdxT, 4>;
+              }
               encoder.add_kernel_node(
-                  cu::copy_gg<InType, OutType, IdxT>,
+                  kernel,
-                  num_blocks,
+                  {num_blocks_x, num_blocks_y},
                   block_dims,
                   0,
                   in_ptr,
                   out_ptr,
-                  data_size,
+                  rest,
                   const_param(shape),
                   const_param(strides_in),
                   const_param(strides_out),

mlx/backend/cuda/copy/copy_general_input.cu

@@ -10,33 +10,67 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT, int NDIM>
+template <typename In, typename Out, typename IdxT, int NDIM, int N_READS>
 __global__ void copy_g_nd(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
-    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_in) {
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides) {
-  IdxT index = cg::this_grid().thread_rank();
+  auto block = cg::this_thread_block();
-  if (index < size) {
+  auto grid = cg::this_grid();
-    IdxT idx_in = elem_to_loc_nd<NDIM>(index, shape.data(), strides_in.data());
+  IdxT index_rest =
-    out[index] = CastOp<In, Out>{}(in[idx_in]);
+      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
+  if (index_rest >= size_rest) {
+    return;
   }
+
+  auto shape_x = shape[NDIM - 1];
+  auto stride_x = strides[NDIM - 1];
+  IdxT index_x =
+      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  auto idx =
+      elem_to_loc_nd<NDIM>(index_rest * shape_x, shape.data(), strides.data());
+  auto in_vec =
+      load_vector<N_READS>(in + idx, index_x, shape_x, stride_x, In(0));
+  AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    out_vec[i] = CastOp<In, Out>{}(in_vec[i]);
+  }
+  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
-template <typename In, typename Out, typename IdxT>
+template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_g(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ Shape shape,
-    const __grid_constant__ Strides strides_in,
+    const __grid_constant__ Strides strides,
     int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
+  auto block = cg::this_thread_block();
-  if (index < size) {
+  auto grid = cg::this_grid();
-    IdxT idx_in = elem_to_loc(index, shape.data(), strides_in.data(), ndim);
+  IdxT index_rest =
-    out[index] = CastOp<In, Out>{}(in[idx_in]);
+      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
+  if (index_rest >= size_rest) {
+    return;
   }
+
+  auto shape_x = shape[ndim - 1];
+  auto stride_x = strides[ndim - 1];
+  IdxT index_x =
+      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  auto idx =
+      elem_to_loc(index_rest * shape_x, shape.data(), strides.data(), ndim);
+  auto in_vec =
+      load_vector<N_READS>(in + idx, index_x, shape_x, stride_x, In(0));
+  AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    out_vec[i] = CastOp<In, Out>{}(in_vec[i]);
+  }
+  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
 } // namespace cu
@@ -61,30 +95,49 @@ void copy_general_input(
             const InType* in_ptr = in.data<InType>() + offset_in;
             OutType* out_ptr = out.data<OutType>() + offset_out;
             int ndim = shape.size();
+            int work_per_thread = 1;
+            auto dim0 = ndim > 0 ? shape.back() : 1;
+            auto rest = out.size() / dim0;
+            if (dim0 >= 4) {
+              work_per_thread = 4;
+            }
+            dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
+            auto block_dims = get_block_dims(dim0, rest, 1);
+            uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
+            uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
+
             if (ndim <= 3) {
               dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                auto [num_blocks, block_dims] = get_launch_args(out, large());
+                auto kernel =
+                    cu::copy_g_nd<InType, OutType, IdxT, dims_constant(), 1>;
+                if (work_per_thread == 4) {
+                  kernel =
+                      cu::copy_g_nd<InType, OutType, IdxT, dims_constant(), 4>;
+                }
                 encoder.add_kernel_node(
-                    cu::copy_g_nd<InType, OutType, IdxT, dims_constant()>,
+                    kernel,
-                    num_blocks,
+                    {num_blocks_x, num_blocks_y},
                     block_dims,
                     0,
                     in_ptr,
                     out_ptr,
-                    out.size(),
+                    rest,
                     const_param<dims_constant()>(shape),
                     const_param<dims_constant()>(strides_in));
               });
             } else { // ndim >= 4
-              auto [num_blocks, block_dims] = get_launch_args(out, large());
+              auto kernel = cu::copy_g<InType, OutType, IdxT, 1>;
+              if (work_per_thread == 4) {
+                kernel = cu::copy_g<InType, OutType, IdxT, 4>;
+              }
               encoder.add_kernel_node(
-                  cu::copy_g<InType, OutType, IdxT>,
+                  kernel,
-                  num_blocks,
+                  {num_blocks_x, num_blocks_y},
                   block_dims,
                   0,
                   in_ptr,
                   out_ptr,
-                  out.size(),
+                  rest,
                   const_param(shape),
                   const_param(strides_in),
                   ndim);

mlx/backend/cuda/cudnn_utils.cpp

@@ -0,0 +1,272 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/cudnn_utils.h"
+#include "mlx/backend/cuda/device.h"
+
+namespace mlx::core {
+
+namespace {
+
+// Create a cudnn tensor descriptor.
+template <typename Vec>
+inline cudnn_frontend::Tensor build_cudnn_tensor(
+    int64_t id,
+    const array& x,
+    const Vec& shape,
+    const Vec& strides) {
+  return cudnn_frontend::TensorBuilder()
+      .setDim(shape.size(), shape.data())
+      .setStrides(strides.size(), strides.data())
+      .setId(id)
+      .setAlignment(get_alignment(x))
+      .setDataType(dtype_to_cudnn_type(x.dtype()))
+      .build();
+}
+
+// In MLX a singleton dim (shape[dim] == 1) can have any stride, but in cuDNN
+// whether a tensor is contiguous is determined with:
+// shape[dim] == shape[dim + 1] * strides[dim + 1]
+// So a contiguous array with singleton dims in MLX may be mistakenly treated
+// as strided in cuDNN, and we work around it by normalizing the strides.
+Strides normalized_strides(const array& x) {
+  if (!x.flags().row_contiguous || x.ndim() < 2) {
+    return x.strides();
+  }
+  Strides strides = x.strides();
+  for (int i = x.ndim() - 2; i >= 0; --i) {
+    if (x.shape(i) == 1) {
+      strides[i] = x.shape(i + 1) * strides[i + 1];
+    }
+  }
+  return strides;
+}
+
+// Return the shape and strides after transposing from NHWC to NCHW.
+auto nhwc_to_nchw(SmallVector<int64_t> shape, SmallVector<int64_t> strides) {
+  assert(shape.size() >= 3);
+  shape.insert(shape.begin() + 1, shape.back());
+  shape.erase(shape.end() - 1);
+  strides.insert(strides.begin() + 1, strides.back());
+  strides.erase(strides.end() - 1);
+  return std::make_tuple(std::move(shape), std::move(strides));
+}
+
+inline auto nhwc_to_nchw(const array& x) {
+  return nhwc_to_nchw(
+      convert_vector<int64_t>(x.shape()), normalized_strides(x));
+}
+
+// Return available engines for a |op_graph|.
+cudnn_frontend::EngineConfigList get_cudnn_engine_configs(
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    cudnn_frontend::OperationGraph& op_graph,
+    bool use_fallback = true) {
+  SmallVector<cudnn_frontend::GeneratorSource, 2> sources;
+  sources.push_back([](auto& op_graph) {
+    auto heuristics = cudnn_frontend::EngineHeuristicsBuilder()
+                          .setOperationGraph(op_graph)
+                          .setHeurMode(CUDNN_HEUR_MODE_A)
+                          .build();
+    return heuristics.getEngineConfig(heuristics.getEngineConfigCount());
+  });
+  if (use_fallback) {
+    sources.push_back([&backend_type](auto& op_graph) {
+      auto fallback = cudnn_frontend::EngineFallbackListBuilder()
+                          .setOperationGraph(op_graph)
+                          .setOperation(backend_type)
+                          .build();
+      return fallback.getFallbackList();
+    });
+  }
+
+  auto configs =
+      cudnn_frontend::EngineConfigGenerator(sources.size(), sources.data())
+          .generate_engine_config(op_graph);
+
+  cudnn_frontend::EngineConfigList filtered_configs;
+  cudnn_frontend::filter(configs, filtered_configs, [dtype](auto c) {
+    if (cudnn_frontend::hasNumericalNote<
+            CUDNN_NUMERICAL_NOTE_DOWN_CONVERT_INPUTS>(c)) {
+      return true;
+    }
+    if (cudnn_frontend::hasNumericalNote<CUDNN_NUMERICAL_NOTE_TENSOR_CORE>(c) &&
+        dtype == float32 && !env::enable_tf32()) {
+      return true;
+    }
+    return false;
+  });
+  return filtered_configs;
+}
+
+// Take |engine_configs| and |op_graph| and find a working execution plans
+// from them.
+std::optional<cudnn_frontend::ExecutionPlan>
+find_cudnn_plan_from_engine_configs(
+    cudnnHandle_t handle,
+    const cudnn_frontend::EngineConfigList& engine_configs,
+    const cudnn_frontend::OperationGraph& op_graph) {
+  auto op_graph_tag = op_graph.getTag();
+  for (const auto& config : engine_configs) {
+    try {
+      return cudnn_frontend::ExecutionPlanBuilder()
+          .setHandle(handle)
+          .setEngineConfig(config, op_graph_tag)
+          .build();
+    } catch (cudnn_frontend::cudnnException& error) {
+      if (error.getCudnnStatus() != CUDNN_STATUS_NOT_SUPPORTED) {
+        throw;
+      }
+    }
+  }
+  return std::nullopt;
+}
+
+// Prepare workspace and args to execute plan.
+template <typename F>
+bool prepare_cudnn_plan(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs,
+    F&& execute) {
+  int workspace_size = plan.getWorkspaceSize();
+  array workspace(
+      workspace_size > 0 ? allocator::malloc(workspace_size)
+                         : allocator::Buffer(nullptr),
+      {workspace_size},
+      uint8);
+
+  auto args = cudnn_frontend::VariantPackBuilder()
+                  .setWorkspacePointer(workspace.data<void>())
+                  .setDataPointers(num_args, data_ptrs)
+                  .setUids(num_args, uids)
+                  .build();
+
+  auto handle = encoder.device().cudnn_handle();
+  cudnnSetStream(handle, encoder.stream());
+
+  if (!execute(handle, plan.get_raw_desc(), args.get_raw_desc())) {
+    return false;
+  }
+
+  encoder.add_temporary(workspace);
+  return true;
+}
+
+} // namespace
+
+cudnn_frontend::Tensor build_cudnn_tensor(int64_t id, const array& x) {
+  auto shape = convert_vector<int64_t>(x.shape());
+  return build_cudnn_tensor(id, x, shape, normalized_strides(x));
+}
+
+cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x) {
+  auto [shape, strides] = nhwc_to_nchw(x);
+  return build_cudnn_tensor(id, x, shape, strides);
+}
+
+cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x) {
+  if (x.ndim() == 0) {
+    SmallVector<int64_t, 4> scalar_dims = {1, 1, 1, 1};
+    return build_cudnn_tensor(id, x, scalar_dims, scalar_dims);
+  }
+  if (x.ndim() == 1) {
+    int64_t s = x.shape(0);
+    SmallVector<int64_t, 4> shape = {1, x.shape(0), 1, 1};
+    SmallVector<int64_t, 4> strides = {s, 1, s, s};
+    return build_cudnn_tensor(id, x, shape, strides);
+  }
+  if (x.ndim() == 2) {
+    int64_t s =
+        x.flags().row_contiguous ? x.shape(1) * x.strides(1) : x.strides(0);
+    SmallVector<int64_t, 4> shape = {x.shape(0), x.shape(1), 1, 1};
+    SmallVector<int64_t, 4> strides = {s, x.strides(1), s, s};
+    return build_cudnn_tensor(id, x, shape, strides);
+  }
+  if (x.ndim() == 3 || x.ndim() == 4) {
+    return build_cudnn_tensor_nchw(id, x);
+  }
+  throw std::runtime_error(
+      fmt::format("Unsupported array with {} dims.", x.ndim()));
+}
+
+cudnn_frontend::Tensor build_cudnn_scalar_4d(int64_t id, Dtype dtype) {
+  SmallVector<int64_t, 4> scalar_dims = {1, 1, 1, 1};
+  return cudnn_frontend::TensorBuilder()
+      .setDim(scalar_dims.size(), scalar_dims.data())
+      .setStrides(scalar_dims.size(), scalar_dims.data())
+      .setId(id)
+      .setAlignment(16)
+      .setDataType(dtype_to_cudnn_type(dtype))
+      .setByValue(true)
+      .build();
+}
+
+std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
+    cudnnHandle_t handle,
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    cudnn_frontend::OperationGraph& op_graph) {
+  auto engine_configs = get_cudnn_engine_configs(backend_type, dtype, op_graph);
+  return find_cudnn_plan_from_engine_configs(handle, engine_configs, op_graph);
+}
+
+bool encode_cudnn_plan_with_capturing(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs) {
+  return prepare_cudnn_plan(
+      encoder,
+      plan,
+      num_args,
+      uids,
+      data_ptrs,
+      [&](auto handle, auto plan, auto args) {
+        auto capture = encoder.capture_context();
+        if (cudnnBackendExecute(handle, plan, args) != CUDNN_STATUS_SUCCESS) {
+          // Discard the captured graph when failed.
+          capture.discard = true;
+          return false;
+        }
+        return true;
+      });
+}
+
+#if CUDNN_VERSION >= 90500
+bool encode_cudnn_plan_with_graph_api(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    CudaGraph& graph,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs) {
+  return prepare_cudnn_plan(
+      encoder,
+      plan,
+      num_args,
+      uids,
+      data_ptrs,
+      [&](auto handle, auto plan, auto args) {
+        if (!graph) {
+          graph = CudaGraph(encoder.device());
+          if (cudnnBackendPopulateCudaGraph(handle, plan, args, graph) !=
+              CUDNN_STATUS_SUCCESS) {
+            return false;
+          }
+        } else {
+          if (cudnnBackendUpdateCudaGraph(handle, plan, args, graph) !=
+              CUDNN_STATUS_SUCCESS) {
+            return false;
+          }
+        }
+        encoder.add_graph_node(graph);
+        return true;
+      });
+}
+#endif
+
+} // namespace mlx::core

mlx/backend/cuda/cudnn_utils.h

@@ -0,0 +1,164 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+#include "mlx/backend/cuda/device/config.h"
+#include "mlx/backend/cuda/utils.h"
+#include "mlx/dtype_utils.h"
+
+#include <cudnn_frontend.h>
+#include <cudnn_frontend_find_plan.h>
+#include <fmt/format.h>
+
+#include <algorithm>
+#include <array>
+
+namespace mlx::core {
+
+namespace cu {
+class CommandEncoder;
+}
+
+// Return pointer alignment of |x|'s data.
+inline uint8_t get_alignment(const array& x) {
+  uint8_t alignment = 1;
+  uintptr_t address = reinterpret_cast<uintptr_t>(x.data<void>());
+  for (; alignment < 32; alignment *= 2) {
+    if (address % (alignment * 2)) {
+      return alignment;
+    }
+  }
+  return alignment;
+}
+
+// Convert the type of elements in |vec| to |T|.
+template <typename T, typename Vec>
+inline SmallVector<T> convert_vector(const Vec& vec) {
+  return SmallVector<T>(vec.begin(), vec.end());
+}
+
+// Return an array that can be used as map key for |vec| with size <= MAX_NDIM.
+//
+// There are 2 differences from the const_param util from kernel_utils.cuh:
+// 1. The rest of array is filled with 0.
+// 2. This util can be used in .cpp files.
+template <typename T, template <typename U> class Vec>
+inline std::array<T, MAX_NDIM> vector_key(const Vec<T>& vec) {
+  if (vec.size() > MAX_NDIM) {
+    throw std::runtime_error(
+        fmt::format("ndim can not be larger than {}.", MAX_NDIM));
+  }
+  std::array<T, MAX_NDIM> result = {};
+  std::copy_n(vec.begin(), vec.size(), result.begin());
+  return result;
+}
+
+// Helpers used by get_data_ptrs to get pointers.
+inline void* get_data_ptr(const array& arr) {
+  return const_cast<void*>(arr.data<void>());
+}
+
+template <typename T, typename = std::enable_if_t<std::is_scalar_v<T>>>
+inline void* get_data_ptr(T& scalar) {
+  return &scalar;
+}
+
+// Return an array filled with data pointers of args.
+template <typename... Args>
+inline std::array<void*, sizeof...(Args)> get_data_ptrs(Args&... args) {
+  return {get_data_ptr(args)...};
+}
+
+// Map dtype to cudnn data type.
+inline cudnnDataType_t dtype_to_cudnn_type(Dtype dtype) {
+  switch (dtype) {
+    case int8:
+      return CUDNN_DATA_INT8;
+    case int32:
+      return CUDNN_DATA_INT32;
+    case uint8:
+      return CUDNN_DATA_UINT8;
+    case float16:
+      return CUDNN_DATA_HALF;
+    case bfloat16:
+      return CUDNN_DATA_BFLOAT16;
+    case float32:
+      return CUDNN_DATA_FLOAT;
+    case float64:
+      return CUDNN_DATA_DOUBLE;
+    default:
+      throw std::runtime_error(fmt::format(
+          "Unsupported dtype in Convolution: {}.", dtype_to_string(dtype)));
+  }
+}
+
+// Create a tensor descriptor from |x|.
+cudnn_frontend::Tensor build_cudnn_tensor(int64_t id, const array& x);
+
+// Create a tensor descriptor from |x|, and transpose from NHWC to NCHW.
+cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x);
+
+// Create a tensor descriptor from |x|, make sure it is 4D, and transpose it
+// from NHWC to NCHW.
+cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x);
+
+// Create a 4D scalar tensor descriptor, which is passed by value.
+cudnn_frontend::Tensor build_cudnn_scalar_4d(int64_t id, Dtype dtype);
+
+// Find a working plan for |op_graph|.
+std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
+    cudnnHandle_t handle,
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    cudnn_frontend::OperationGraph& op_graph);
+
+// Encode the plan to command buffer by capturing.
+bool encode_cudnn_plan_with_capturing(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs);
+
+#if CUDNN_VERSION >= 90500
+// Encode the plan to command buffer by using native graph api of cudnn. If the
+// |graph| is empty it will be populated, otherwise it will be updated.
+bool encode_cudnn_plan_with_graph_api(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    CudaGraph& graph,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs);
+#endif
+
+// Helpers to make calls like encode_cudnn_plan(..., {'x', 'y', 'z'}, x, y, z).
+template <typename... Args>
+bool encode_cudnn_plan(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    std::initializer_list<int64_t> uids,
+    Args&... args) {
+  assert(uids.size() == sizeof...(args));
+  auto data_ptrs = get_data_ptrs(args...);
+  return encode_cudnn_plan_with_capturing(
+      encoder, plan, uids.size(), uids.begin(), data_ptrs.data());
+}
+
+#if CUDNN_VERSION >= 90500
+template <typename... Args>
+bool encode_cudnn_plan(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    CudaGraph& graph,
+    std::initializer_list<int64_t> uids,
+    Args&... args) {
+  assert(uids.size() == sizeof...(args));
+  auto data_ptrs = get_data_ptrs(args...);
+  return encode_cudnn_plan_with_graph_api(
+      encoder, plan, graph, uids.size(), uids.begin(), data_ptrs.data());
+}
+#endif
+
+} // namespace mlx::core

mlx/backend/cuda/custom_kernel.cpp

@@ -0,0 +1,379 @@
+// Copyright © 2025 Apple Inc.
+
+#include <iostream>
+
+#include "mlx/backend/common/compiled.h"
+#include "mlx/backend/cuda/jit_module.h"
+#include "mlx/backend/cuda/utils.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/fast.h"
+#include "mlx/fast_primitives.h"
+
+#include <fmt/format.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core::fast {
+
+namespace {
+
+constexpr const char* default_header = R"(
+#include "mlx/backend/cuda/device/utils.cuh"
+
+#include <cooperative_groups.h>
+
+#define inf cuda::std::numeric_limits<float>::infinity()
+
+)";
+
+std::string template_arguments_hash(
+    const std::vector<std::pair<std::string, TemplateArg>>& template_args) {
+  if (template_args.empty()) {
+    return "";
+  }
+
+  std::string hash;
+  hash.reserve(512);
+
+  for (const auto& [name, arg] : template_args) {
+    if (std::holds_alternative<int>(arg)) {
+      hash += fmt::format("_{}", std::get<int>(arg));
+    } else if (std::holds_alternative<bool>(arg)) {
+      hash += (std::get<bool>(arg)) ? "_t" : "_f";
+    } else if (std::holds_alternative<Dtype>(arg)) {
+      hash += "_";
+      hash += get_type_string(std::get<Dtype>(arg));
+    }
+  }
+
+  return hash;
+}
+
+std::string build_kernel(
+    const std::string& func_name,
+    const std::string& header,
+    const std::string& source,
+    const std::vector<std::string>& input_names,
+    const std::vector<array>& inputs,
+    const std::vector<std::string>& output_names,
+    const std::vector<Dtype>& output_dtypes,
+    const std::vector<std::pair<std::string, TemplateArg>>& template_args,
+    const std::vector<CustomKernelShapeInfo>& shape_infos) {
+  std::string kernel_source;
+  kernel_source.reserve(header.size() + source.size() + 8192);
+  kernel_source += default_header;
+  kernel_source += header;
+  kernel_source +=
+      "namespace mlx::core::cu {\n\n"
+      "namespace cg = cooperative_groups;\n\n";
+
+  kernel_source += "__global__ void ";
+  kernel_source += func_name;
+  kernel_source += "(\n";
+
+  // Add inputs
+  for (int i = 0; i < inputs.size(); ++i) {
+    const auto& name = input_names[i];
+    const auto& arr = inputs[i];
+    kernel_source += "    const ";
+    kernel_source += dtype_to_cuda_type(arr.dtype());
+    kernel_source += "* ";
+    kernel_source += name;
+    kernel_source += ",\n";
+    // Add input shape, strides and ndim if present in the source
+    if (arr.ndim() > 0) {
+      if (shape_infos[i].shape) {
+        kernel_source += "    const __grid_constant__ Shape ";
+        kernel_source += name;
+        kernel_source += "_shape,\n";
+      }
+      if (shape_infos[i].strides) {
+        kernel_source += "    const __grid_constant__ Strides ";
+        kernel_source += name;
+        kernel_source += "_strides,\n";
+      }
+      if (shape_infos[i].ndim) {
+        kernel_source += "    const __grid_constant__ int ";
+        kernel_source += name;
+        kernel_source += "_ndim,\n";
+      }
+    }
+  }
+
+  // Add outputs
+  for (int i = 0; i < output_names.size(); ++i) {
+    const auto& name = output_names[i];
+    const auto& dtype = output_dtypes[i];
+    kernel_source += "    ";
+    kernel_source += dtype_to_cuda_type(dtype);
+    kernel_source += "* ";
+    kernel_source += name;
+    if (i < output_names.size() - 1) {
+      kernel_source += ",\n";
+    } else {
+      kernel_source += ") {\n";
+    }
+  }
+
+  // Set compile time constants
+  if (!template_args.empty()) {
+    for (const auto& [name, arg] : template_args) {
+      if (std::holds_alternative<int>(arg)) {
+        kernel_source +=
+            fmt::format("  constexpr int {} = {};\n", name, std::get<int>(arg));
+      } else if (std::holds_alternative<bool>(arg)) {
+        kernel_source += fmt::format(
+            "  constexpr bool {} = {};\n", name, std::get<bool>(arg));
+      } else {
+        kernel_source += fmt::format(
+            "  using {} = {};\n",
+            name,
+            dtype_to_cuda_type(std::get<Dtype>(arg)));
+      }
+    }
+    kernel_source += "\n";
+  }
+
+  kernel_source += source;
+  kernel_source += "\n}\n\n} // namespace mlx::core::cu\n";
+
+  return kernel_source;
+}
+
+} // namespace
+
+CustomKernelFunction cuda_kernel(
+    const std::string& name,
+    const std::vector<std::string>& input_names,
+    const std::vector<std::string>& output_names,
+    const std::string& source,
+    const std::string& header,
+    bool ensure_row_contiguous,
+    int shared_memory) {
+  if (output_names.empty()) {
+    throw std::invalid_argument(
+        "[custom_kernel] Must specify at least one output.");
+  }
+
+  std::vector<CustomKernelShapeInfo> shape_infos;
+  for (auto& n : input_names) {
+    CustomKernelShapeInfo shape_info;
+    shape_info.shape = source.find(n + "_shape") != std::string::npos;
+    shape_info.strides = source.find(n + "_strides") != std::string::npos;
+    shape_info.ndim = source.find(n + "_ndim") != std::string::npos;
+    shape_infos.push_back(shape_info);
+  }
+
+  return [=, shape_infos = std::move(shape_infos)](
+             const std::vector<array>& inputs,
+             const std::vector<Shape>& output_shapes,
+             const std::vector<Dtype>& output_dtypes,
+             std::tuple<int, int, int> grid,
+             std::tuple<int, int, int> threadgroup,
+             const std::vector<std::pair<std::string, TemplateArg>>&
+                 template_args = {},
+             std::optional<float> init_value = std::nullopt,
+             bool verbose = false,
+             StreamOrDevice s_ = {}) {
+    if (inputs.size() != input_names.size()) {
+      std::ostringstream msg;
+      msg << "[custom_kernel] Expected `inputs` to have size "
+          << input_names.size() << " but got size " << inputs.size() << "."
+          << std::endl;
+      throw std::invalid_argument(msg.str());
+    }
+    if (output_shapes.size() != output_names.size()) {
+      std::ostringstream msg;
+      msg << "[custom_kernel] Expected `output_shapes` to have size "
+          << output_names.size() << " but got size " << output_shapes.size()
+          << "." << std::endl;
+      throw std::invalid_argument(msg.str());
+    }
+    if (output_dtypes.size() != output_names.size()) {
+      std::ostringstream msg;
+      msg << "[custom_kernel] Expected `output_dtypes` to have size "
+          << output_names.size() << " but got size " << output_dtypes.size()
+          << "." << std::endl;
+      throw std::invalid_argument(msg.str());
+    }
+
+    auto s = to_stream(s_);
+    if (s.device != Device::gpu) {
+      throw std::invalid_argument("[custom_kernel] Only supports the GPU.");
+    }
+
+    std::string kernel_name =
+        "custom_kernel_" + name + template_arguments_hash(template_args);
+    std::string kernel_source = build_kernel(
+        kernel_name,
+        header,
+        source,
+        input_names,
+        inputs,
+        output_names,
+        output_dtypes,
+        template_args,
+        shape_infos);
+
+    if (verbose) {
+      std::cout << "Generated source code for `" << kernel_name
+                << "`:" << std::endl
+                << "```" << std::endl
+                << kernel_source << std::endl
+                << "```" << std::endl;
+    }
+
+    return array::make_arrays(
+        std::move(output_shapes),
+        std::move(output_dtypes),
+        std::make_shared<CustomKernel>(
+            s,
+            std::move(kernel_name),
+            std::move(kernel_source),
+            grid,
+            threadgroup,
+            shape_infos,
+            ensure_row_contiguous,
+            init_value,
+            std::vector<ScalarArg>{},
+            false,
+            shared_memory),
+        std::move(inputs));
+  };
+}
+
+std::vector<array> precompiled_cuda_kernel(
+    const std::string& name,
+    const std::string& compiled_source,
+    const std::vector<array>& inputs,
+    const std::vector<Shape>& output_shapes,
+    const std::vector<Dtype>& output_dtypes,
+    const std::vector<ScalarArg>& scalars,
+    std::tuple<int, int, int> grid,
+    std::tuple<int, int, int> threadgroup,
+    int shared_memory,
+    std::optional<float> init_value,
+    bool ensure_row_contiguous,
+    StreamOrDevice s) {
+  std::vector<CustomKernelShapeInfo> shape_infos(
+      inputs.size(), CustomKernelShapeInfo{false, false, false});
+  return array::make_arrays(
+      output_shapes,
+      output_dtypes,
+      std::make_shared<CustomKernel>(
+          to_stream(s),
+          name,
+          compiled_source,
+          grid,
+          threadgroup,
+          shape_infos,
+          ensure_row_contiguous,
+          init_value,
+          scalars,
+          true,
+          shared_memory),
+      inputs);
+}
+
+void CustomKernel::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  nvtx3::scoped_range r("CustomKernel::eval_gpu");
+  auto& s = stream();
+
+  std::vector<array> copies;
+
+  // Allocate and initialize the output arrays
+  for (auto& out : outputs) {
+    if (init_value_) {
+      copies.emplace_back(init_value_.value(), out.dtype());
+      fill_gpu(copies.back(), out, s);
+    } else {
+      out.set_data(allocator::malloc(out.nbytes()));
+    }
+  }
+
+  // Create the input arrays and copy if needed
+  auto check_input = [&copies, &s, this](const array& x) -> const array {
+    bool no_copy = x.flags().row_contiguous;
+    if (!ensure_row_contiguous_ || no_copy) {
+      return x;
+    } else {
+      copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
+      copy_gpu(x, copies.back(), CopyType::General, s);
+      return copies.back();
+    }
+  };
+  std::vector<array> checked_inputs;
+  for (const array& in : inputs) {
+    checked_inputs.push_back(check_input(in));
+  }
+
+  // Compile the custom kernel
+  std::string kernel_name =
+      (is_precompiled_) ? name_ : "mlx::core::cu::" + name_;
+  cu::JitModule& mod = cu::get_jit_module(
+      s.device,
+      name_,
+      [&]() {
+        return std::make_tuple(
+            is_precompiled_, source_, std::vector{kernel_name});
+      },
+      false);
+
+  // Make the arguments
+  cu::KernelArgs args;
+  for (int i = 0; i < checked_inputs.size(); i++) {
+    const array& in = checked_inputs[i];
+    auto& shape_info = shape_infos_[i];
+    args.append(in);
+    if (shape_info.shape) {
+      args.append_ndim(in.shape());
+    }
+    if (shape_info.strides) {
+      args.append_ndim(in.strides());
+    }
+    if (shape_info.ndim) {
+      args.append<int32_t>(in.ndim());
+    }
+  }
+  for (auto& out : outputs) {
+    args.append(out);
+  }
+  for (auto& s : scalar_arguments_) {
+    if (std::holds_alternative<bool>(s)) {
+      args.append(std::get<bool>(s));
+    } else if (std::holds_alternative<int>(s)) {
+      args.append(std::get<int>(s));
+    } else if (std::holds_alternative<float>(s)) {
+      args.append(std::get<float>(s));
+    }
+  }
+
+  // Make the grid
+  const auto [tx, ty, tz] = threadgroup_;
+  const auto [gx, gy, gz] = grid_;
+  dim3 block(std::min(tx, gx), std::min(ty, gy), std::min(tz, gz));
+  dim3 grid((gx + tx - 1) / tx, (gy + ty - 1) / ty, (gz + tz - 1) / tz);
+
+  // Call the kernel
+  auto& encoder = cu::get_command_encoder(s);
+  for (const auto& in : checked_inputs) {
+    encoder.set_input_array(in);
+  }
+  for (const auto& out : outputs) {
+    encoder.set_output_array(out);
+  }
+  for (const auto& t : copies) {
+    encoder.add_temporary(t);
+  }
+  auto kernel =
+      mod.get_kernel(kernel_name, [smem = shared_memory_](CUfunction kernel) {
+        if (smem > 0 && smem > 48000) {
+          cuFuncSetAttribute(
+              kernel, CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES, smem);
+        }
+      });
+  encoder.add_kernel_node(kernel, grid, block, shared_memory_, args.args());
+}
+
+} // namespace mlx::core::fast

mlx/backend/cuda/device.cpp

@@ -27,11 +27,11 @@ void check_cudnn_error(const char* name, cudnnStatus_t err) {
   }
 }
 
-int cuda_graph_cache_size() {
+bool use_cuda_graphs() {
-  static int cache_size = []() {
+  static bool use_graphs = []() {
-    return env::get_var("MLX_CUDA_GRAPH_CACHE_SIZE", 100);
+    return env::get_var("MLX_USE_CUDA_GRAPHS", true);
   }();
-  return cache_size;
+  return use_graphs;
 }
 
 } // namespace
@@ -86,14 +86,19 @@ CommandEncoder& Device::get_command_encoder(Stream s) {
 
 CommandEncoder::CaptureContext::CaptureContext(CommandEncoder& enc) : enc(enc) {
   enc.device().make_current();
+  if (!use_cuda_graphs()) {
+    return;
+  }
   CHECK_CUDA_ERROR(
       cudaStreamBeginCapture(enc.stream(), cudaStreamCaptureModeGlobal));
 }
 
 CommandEncoder::CaptureContext::~CaptureContext() {
-  CHECK_CUDA_ERROR(cudaStreamEndCapture(enc.stream(), &graph));
+  if (!use_cuda_graphs()) {
-  std::unique_ptr<cudaGraph_t, void (*)(cudaGraph_t*)> graph_freer(
+    return;
-      &graph, [](cudaGraph_t* p) { CHECK_CUDA_ERROR(cudaGraphDestroy(*p)); });
+  }
+
+  graph.end_capture(enc.stream());
   if (discard) {
     return;
   }
@@ -107,6 +112,9 @@ CommandEncoder::ConcurrentContext::ConcurrentContext(CommandEncoder& enc)
 
 CommandEncoder::ConcurrentContext::~ConcurrentContext() {
   enc.in_concurrent_ = false;
+  if (!use_cuda_graphs()) {
+    return;
+  }
 
   // Use an empty graph node for synchronization
   CommandEncoder::GraphNode empty{NULL, 'E', std::to_string(enc.node_count_++)};
@@ -185,20 +193,28 @@ void CommandEncoder::insert_graph_dependencies(std::vector<GraphNode> nodes) {
 }
 
 CommandEncoder::CommandEncoder(Device& d)
-    : device_(d), stream_(d), graph_cache_(cuda_graph_cache_size()) {
+    : device_(d),
-  CHECK_CUDA_ERROR(cudaGraphCreate(&graph_, 0));
+      stream_(d),
-}
+      graph_(d),
+      graph_cache_("MLX_CUDA_GRAPH_CACHE_SIZE", /* default_capacity */ 400) {}
 
 void CommandEncoder::add_completed_handler(std::function<void()> task) {
   worker_.add_task(std::move(task));
 }
 
 void CommandEncoder::set_input_array(const array& arr) {
+  if (!use_cuda_graphs()) {
+    return;
+  }
   auto id = reinterpret_cast<std::uintptr_t>(arr.buffer().ptr());
   active_deps_.push_back(id);
 }
 
 void CommandEncoder::set_output_array(const array& arr) {
+  if (!use_cuda_graphs()) {
+    return;
+  }
+
   auto id = reinterpret_cast<std::uintptr_t>(arr.buffer().ptr());
   active_deps_.push_back(id);
   active_outputs_.push_back(id);
@@ -216,6 +232,11 @@ void CommandEncoder::add_kernel_node(
     dim3 block_dim,
     uint32_t smem_bytes,
     void** params) {
+  if (!use_cuda_graphs()) {
+    CHECK_CUDA_ERROR(cudaLaunchKernel(
+        func, grid_dim, block_dim, params, smem_bytes, stream()));
+    return;
+  }
   cudaKernelNodeParams kernel_params = {0};
   kernel_params.func = func;
   kernel_params.gridDim = grid_dim;
@@ -231,6 +252,22 @@ void CommandEncoder::add_kernel_node(
     dim3 block_dim,
     uint32_t smem_bytes,
     void** params) {
+  if (!use_cuda_graphs()) {
+    CHECK_CUDA_ERROR(cuLaunchKernel(
+        func,
+        grid_dim.x,
+        grid_dim.y,
+        grid_dim.z,
+        block_dim.x,
+        block_dim.y,
+        block_dim.z,
+        smem_bytes,
+        stream(),
+        params,
+        nullptr));
+    return;
+  }
+
   CUDA_KERNEL_NODE_PARAMS kernel_params = {0};
   kernel_params.func = func;
   kernel_params.gridDimX = grid_dim.x;
@@ -257,6 +294,13 @@ void CommandEncoder::add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params) {
 }
 
 void CommandEncoder::add_graph_node(cudaGraph_t child) {
+  if (!use_cuda_graphs()) {
+    CudaGraphExec graph_exec;
+    graph_exec.instantiate(child);
+    device_.make_current();
+    CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream()));
+    return;
+  }
   cudaGraphNode_t node;
   CHECK_CUDA_ERROR(cudaGraphAddChildGraphNode(&node, graph_, NULL, 0, child));
   insert_graph_dependencies(GraphNode{node, 'G'});
@@ -270,7 +314,13 @@ void CommandEncoder::commit() {
   if (node_count_ > 0) {
     if (!from_nodes_.empty()) {
       CHECK_CUDA_ERROR(cudaGraphAddDependencies(
-          graph_, from_nodes_.data(), to_nodes_.data(), from_nodes_.size()));
+          graph_,
+          from_nodes_.data(),
+          to_nodes_.data(),
+#if CUDART_VERSION >= 13000
+          nullptr, // edgeData
+#endif // CUDART_VERSION >= 13000
+          from_nodes_.size()));
     }
 
     graph_key_ += ".";
@@ -311,8 +361,7 @@ void CommandEncoder::commit() {
     to_nodes_.clear();
     graph_key_.clear();
     node_map_.clear();
-    CHECK_CUDA_ERROR(cudaGraphDestroy(graph_));
+    graph_ = CudaGraph(device_);
-    CHECK_CUDA_ERROR(cudaGraphCreate(&graph_, 0));
   }
 
   // Put completion handlers in a batch.

mlx/backend/cuda/device.h

@@ -21,7 +21,7 @@ class CommandEncoder {
   struct CaptureContext {
     CaptureContext(CommandEncoder& enc);
     ~CaptureContext();
-    cudaGraph_t graph;
+    CudaGraph graph;
     CommandEncoder& enc;
     bool discard{false};
   };
@@ -76,9 +76,6 @@ class CommandEncoder {
       uint32_t smem_bytes,
       void** params);
 
-  // Low-level graph helpers.
-  void add_kernel_node(const cudaKernelNodeParams& params);
-  void add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params);
   void add_graph_node(cudaGraph_t child);
 
   void add_temporary(const array& arr) {
@@ -101,6 +98,9 @@ class CommandEncoder {
   void synchronize();
 
  private:
+  void add_kernel_node(const cudaKernelNodeParams& params);
+  void add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params);
+
   struct GraphNode {
     cudaGraphNode_t node;
     // K = kernel
@@ -115,7 +115,7 @@ class CommandEncoder {
 
   Device& device_;
   CudaStream stream_;
-  cudaGraph_t graph_;
+  CudaGraph graph_;
   Worker worker_;
   char node_count_{0};
   char graph_node_count_{0};

mlx/backend/cuda/device/binary_ops.cuh

@@ -204,6 +204,12 @@ struct Power {
   __device__ T operator()(T base, T exp) {
     if constexpr (cuda::std::is_integral_v<T>) {
       T res = 1;
+      // Raising an integer to a negative power is undefined
+      if constexpr (cuda::std::is_signed_v<T>) {
+        if (exp < 0) {
+          return 0;
+        }
+      }
       while (exp) {
         if (exp & 1) {
           res *= base;

mlx/backend/cuda/device/cast_op.cuh

@@ -6,7 +6,6 @@
 
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
-#include <thrust/iterator/transform_iterator.h>
 
 namespace mlx::core::cu {
 
@@ -116,15 +115,4 @@ inline __host__ __device__ auto cast_to(SrcT x) {
   return CastOp<SrcT, DstT>{}(x);
 }
 
-// Return an iterator that cast the value to DstT using CastOp.
-template <typename DstT, typename Iterator>
-inline __host__ __device__ auto make_cast_iterator(Iterator it) {
-  using SrcT = typename cuda::std::iterator_traits<Iterator>::value_type;
-  if constexpr (std::is_same_v<SrcT, DstT>) {
-    return it;
-  } else {
-    return thrust::make_transform_iterator(it, CastOp<SrcT, DstT>{});
-  }
-}
-
 } // namespace mlx::core::cu

mlx/backend/cuda/device/utils.cuh

@@ -146,6 +146,23 @@ inline __device__ void store_vector(
   }
 }
 
+template <int N, typename T, typename SizeT>
+inline __device__ void store_vector(
+    T* ptr,
+    uint32_t offset,
+    const AlignedVector<T, N>& vec,
+    SizeT size,
+    int64_t stride) {
+  if (is_aligned<N>(ptr) && (offset + 1) * N <= size && stride == 1) {
+    auto* to = reinterpret_cast<AlignedVector<T, N>*>(ptr);
+    to[offset] = vec;
+  } else {
+    for (int i = 0; (offset * N + i) < size && i < N; ++i) {
+      ptr[stride * (offset * N + i)] = vec[i];
+    }
+  }
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // Type limits utils
 ///////////////////////////////////////////////////////////////////////////////

mlx/backend/cuda/distributed.cu

@@ -0,0 +1,56 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/distributed/primitives.h"
+#include "mlx/primitives.h"
+
+#include <cassert>
+
+namespace mlx::core::distributed {
+void AllReduce::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 1);
+  assert(outputs.size() == 1);
+
+  auto set_input_output =
+      [s = stream()](const array& in, array& out) -> std::pair<array, array> {
+    if (!in.flags().row_contiguous) {
+      copy_gpu(in, out, CopyType::General, s);
+      return {out, out};
+    } else if (in.is_donatable()) {
+      out.copy_shared_buffer(in);
+      return {in, out};
+    } else {
+      out.set_data(allocator::malloc(out.nbytes()));
+      return {in, out};
+    }
+  };
+
+  auto [input, output] = set_input_output(inputs[0], outputs[0]);
+
+  auto& encoder = cu::get_command_encoder(stream());
+  encoder.set_input_array(input);
+  encoder.set_output_array(output);
+
+  auto capture = encoder.capture_context();
+  auto& s = stream();
+
+  switch (reduce_type_) {
+    case Sum:
+      distributed::detail::all_sum(group(), input, output, s);
+      break;
+    case Max:
+      distributed::detail::all_max(group(), input, output, s);
+      break;
+    case Min:
+      distributed::detail::all_min(group(), input, output, s);
+      break;
+    default:
+      throw std::runtime_error(
+          "Only all reduce sum, max, and min are supported.");
+  }
+}
+} // namespace mlx::core::distributed

mlx/backend/cuda/eval.cpp

@@ -15,8 +15,9 @@ bool is_available() {
 }
 
 void new_stream(Stream s) {
-  // Force initalization of cuda, so cuda runtime get destroyed at last.
+  // Force initalization of CUDA by creating an event, so the CUDA runtime and
-  cudaFree(nullptr);
+  // our CUDA event pool get destroyed last.
+  cu::CudaEvent(cudaEventDefault);
   // Ensure the static stream objects get created.
   cu::get_command_encoder(s);
 }

mlx/backend/cuda/event.cu

@@ -3,10 +3,12 @@
 #include "mlx/backend/cuda/allocator.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/event.h"
-#include "mlx/backend/cuda/utils.h"
 #include "mlx/event.h"
 #include "mlx/scheduler.h"
 
+#include <map>
+#include <vector>
+
 #include <nvtx3/nvtx3.hpp>
 
 namespace mlx::core {
@@ -17,104 +19,141 @@ namespace cu {
 // CudaEvent implementations
 ///////////////////////////////////////////////////////////////////////////////
 
-// Cuda event managed with RAII.
+namespace {
-class CudaEventHandle {
+
- public:
+// Manage cached cudaEvent_t objects.
-  CudaEventHandle() {
+struct CudaEventPool {
-    CHECK_CUDA_ERROR(cudaEventCreateWithFlags(
+  static CudaEventHandle create(int flags) {
-        &event_, cudaEventDisableTiming | cudaEventBlockingSync));
+    auto& cache = cache_for(flags);
+    if (cache.empty()) {
+      return CudaEventHandle(flags);
+    } else {
+      CudaEventHandle ret = std::move(cache.back());
+      cache.pop_back();
+      return ret;
+    }
   }
 
-  ~CudaEventHandle() {
+  static void release(CudaEventHandle event) {
-    CHECK_CUDA_ERROR(cudaEventDestroy(event_));
+    cache_for(event.flags).push_back(std::move(event));
   }
 
-  CudaEventHandle(const CudaEventHandle&) = delete;
+  static std::vector<CudaEventHandle>& cache_for(int flags) {
-  CudaEventHandle& operator=(const CudaEventHandle&) = delete;
+    static std::map<int, std::vector<CudaEventHandle>> cache;
-
+    return cache[flags];
-  operator cudaEvent_t() const {
-    return event_;
   }
-
- private:
-  cudaEvent_t event_;
 };
 
-CudaEvent::CudaEvent() : event_(std::make_shared<CudaEventHandle>()) {}
+} // namespace
+
+CudaEventHandle::CudaEventHandle(int flags) : flags(flags) {
+  CHECK_CUDA_ERROR(cudaEventCreateWithFlags(&handle_, flags));
+  assert(handle_ != nullptr);
+}
+
+CudaEvent::CudaEvent(int flags) : event_(CudaEventPool::create(flags)) {}
+
+CudaEvent::~CudaEvent() {
+  CudaEventPool::release(std::move(event_));
+}
 
 void CudaEvent::wait() {
   nvtx3::scoped_range r("cu::CudaEvent::wait");
-  if (!recorded_) {
+  cudaEventSynchronize(event_);
-    throw std::runtime_error("Should not wait on a CudaEvent before record.");
-  }
-  cudaEventSynchronize(*event_);
 }
 
 void CudaEvent::wait(cudaStream_t stream) {
-  if (!recorded_) {
+  cudaStreamWaitEvent(stream, event_);
-    throw std::runtime_error("Should not wait on a CudaEvent before record.");
-  }
-  cudaStreamWaitEvent(stream, *event_);
-}
-
-void CudaEvent::wait(Stream s) {
-  if (s.device == mlx::core::Device::cpu) {
-    scheduler::enqueue(s, [*this]() mutable { wait(); });
-  } else {
-    auto& enc = cu::get_command_encoder(s);
-    enc.commit();
-    wait(enc.stream());
-  }
 }
 
 void CudaEvent::record(cudaStream_t stream) {
-  cudaEventRecord(*event_, stream);
+  cudaEventRecord(event_, stream);
-  recorded_ = true;
-}
-
-void CudaEvent::record(Stream s) {
-  if (s.device == mlx::core::Device::cpu) {
-    throw std::runtime_error("CudaEvent can not wait on cpu stream.");
-  } else {
-    auto& enc = cu::get_command_encoder(s);
-    enc.commit();
-    record(enc.stream());
-  }
 }
 
 bool CudaEvent::completed() const {
-  return cudaEventQuery(*event_) == cudaSuccess;
+  return cudaEventQuery(event_) == cudaSuccess;
 }
 
+// Wraps CudaEvent with a few features:
+// 1. The class can be copied.
+// 2. Make wait/record work with CPU streams.
+// 3. Add checks for waiting on un-recorded event.
+class CopyableCudaEvent {
+ public:
+  CopyableCudaEvent()
+      : event_(std::make_shared<CudaEvent>(
+            cudaEventDisableTiming | cudaEventBlockingSync)) {}
+
+  void wait() {
+    event_->wait();
+  }
+
+  void wait(Stream s) {
+    if (s.device == mlx::core::Device::cpu) {
+      scheduler::enqueue(s, [*this]() mutable {
+        check_recorded();
+        event_->wait();
+      });
+    } else {
+      check_recorded();
+      auto& encoder = cu::get_command_encoder(s);
+      encoder.commit();
+      event_->wait(encoder.stream());
+    }
+  }
+
+  void record(Stream s) {
+    if (s.device == mlx::core::Device::cpu) {
+      throw std::runtime_error("CudaEvent can not wait on CPU stream.");
+    } else {
+      auto& encoder = cu::get_command_encoder(s);
+      encoder.commit();
+      event_->record(encoder.stream());
+      recorded_ = true;
+    }
+  }
+
+  bool is_signaled() const {
+    return recorded_ && event_->completed();
+  }
+
+ private:
+  void check_recorded() const {
+    if (!recorded_) {
+      throw std::runtime_error(
+          "Should not wait on a CudaEvent before recording.");
+    }
+  }
+
+  std::shared_ptr<CudaEvent> event_;
+  bool recorded_{false};
+};
+
 ///////////////////////////////////////////////////////////////////////////////
-// SharedEvent implementations
+// AtomicEvent implementations
 ///////////////////////////////////////////////////////////////////////////////
 
-__host__ __device__ void event_wait(SharedEvent::Atomic* ac, uint64_t value) {
+__host__ __device__ void event_wait(AtomicEvent::Atomic* ac, uint64_t value) {
   uint64_t current;
   while ((current = ac->load()) < value) {
     ac->wait(current);
   }
 }
 
-__host__ __device__ void event_signal(SharedEvent::Atomic* ac, uint64_t value) {
+__host__ __device__ void event_signal(AtomicEvent::Atomic* ac, uint64_t value) {
   ac->store(value);
   ac->notify_all();
 }
 
-__global__ void event_wait_kernel(SharedEvent::Atomic* ac, uint64_t value) {
+__global__ void event_wait_kernel(AtomicEvent::Atomic* ac, uint64_t value) {
   event_wait(ac, value);
 }
 
-__global__ void event_signal_kernel(SharedEvent::Atomic* ac, uint64_t value) {
+__global__ void event_signal_kernel(AtomicEvent::Atomic* ac, uint64_t value) {
   event_signal(ac, value);
 }
 
-SharedEvent::Atomic* to_atomic(std::shared_ptr<Buffer> buf) {
+AtomicEvent::AtomicEvent() {
-  return static_cast<SharedEvent::Atomic*>(buf->raw_ptr());
-}
-
-SharedEvent::SharedEvent() {
   buf_ = std::shared_ptr<Buffer>(
       new Buffer{allocator().malloc(sizeof(Atomic))}, [](Buffer* ptr) {
         allocator().free(*ptr);
@@ -123,17 +162,17 @@ SharedEvent::SharedEvent() {
   *static_cast<uint64_t*>(buf_->raw_ptr()) = 0;
 }
 
-void SharedEvent::wait(uint64_t value) {
+void AtomicEvent::wait(uint64_t value) {
-  nvtx3::scoped_range r("cu::SharedEvent::wait");
+  nvtx3::scoped_range r("cu::AtomicEvent::wait");
-  event_wait(to_atomic(buf_), value);
+  event_wait(atomic(), value);
 }
 
-void SharedEvent::wait(cudaStream_t stream, uint64_t value) {
+void AtomicEvent::wait(cudaStream_t stream, uint64_t value) {
-  event_wait_kernel<<<1, 1, 0, stream>>>(to_atomic(buf_), value);
+  event_wait_kernel<<<1, 1, 0, stream>>>(atomic(), value);
 }
 
-void SharedEvent::wait(Stream s, uint64_t value) {
+void AtomicEvent::wait(Stream s, uint64_t value) {
-  nvtx3::scoped_range r("cu::SharedEvent::wait(s)");
+  nvtx3::scoped_range r("cu::AtomicEvent::wait(s)");
   if (s.device == mlx::core::Device::cpu) {
     scheduler::enqueue(s, [*this, value]() mutable { wait(value); });
   } else {
@@ -144,17 +183,17 @@ void SharedEvent::wait(Stream s, uint64_t value) {
   }
 }
 
-void SharedEvent::signal(uint64_t value) {
+void AtomicEvent::signal(uint64_t value) {
-  nvtx3::scoped_range r("cu::SharedEvent::signal");
+  nvtx3::scoped_range r("cu::AtomicEvent::signal");
-  event_signal(to_atomic(buf_), value);
+  event_signal(atomic(), value);
 }
 
-void SharedEvent::signal(cudaStream_t stream, uint64_t value) {
+void AtomicEvent::signal(cudaStream_t stream, uint64_t value) {
-  event_signal_kernel<<<1, 1, 0, stream>>>(to_atomic(buf_), value);
+  event_signal_kernel<<<1, 1, 0, stream>>>(atomic(), value);
 }
 
-void SharedEvent::signal(Stream s, uint64_t value) {
+void AtomicEvent::signal(Stream s, uint64_t value) {
-  nvtx3::scoped_range r("cu::SharedEvent::signal(s)");
+  nvtx3::scoped_range r("cu::AtomicEvent::signal(s)");
   if (s.device == mlx::core::Device::cpu) {
     // Signal through a GPU stream so the atomic is updated in GPU - updating
     // the atomic in CPU sometimes does not get GPU notified.
@@ -168,14 +207,14 @@ void SharedEvent::signal(Stream s, uint64_t value) {
   }
 }
 
-bool SharedEvent::is_signaled(uint64_t value) const {
+bool AtomicEvent::is_signaled(uint64_t value) const {
-  nvtx3::scoped_range r("cu::SharedEvent::is_signaled");
+  nvtx3::scoped_range r("cu::AtomicEvent::is_signaled");
-  return to_atomic(buf_)->load() >= value;
+  return atomic()->load() >= value;
 }
 
-uint64_t SharedEvent::value() const {
+uint64_t AtomicEvent::value() const {
-  nvtx3::scoped_range r("cu::SharedEvent::value");
+  nvtx3::scoped_range r("cu::AtomicEvent::value");
-  return to_atomic(buf_)->load();
+  return atomic()->load();
 }
 
 } // namespace cu
@@ -188,14 +227,14 @@ namespace {
 
 struct EventImpl {
   // CudaEvent is preferred when possible because it is fast, however we have
-  // to fallback to SharedEvent in following cases:
+  // to fallback to AtomicEvent in following cases:
   // 1. the event is used to wait/signal a cpu stream;
   // 2. signal value other than 1 has been specified.
-  std::unique_ptr<cu::CudaEvent> cuda;
+  std::unique_ptr<cu::CopyableCudaEvent> cuda;
-  std::unique_ptr<cu::SharedEvent> shared;
+  std::unique_ptr<cu::AtomicEvent> atomic;
 
   bool is_created() const {
-    return cuda || shared;
+    return cuda || atomic;
   }
 
   void ensure_created(Stream s, uint64_t signal_value) {
@@ -203,10 +242,10 @@ struct EventImpl {
       return;
     }
     if (s.device == mlx::core::Device::cpu || signal_value > 1) {
-      nvtx3::mark("Using slow SharedEvent");
+      nvtx3::mark("Using slow AtomicEvent");
-      shared = std::make_unique<cu::SharedEvent>();
+      atomic = std::make_unique<cu::AtomicEvent>();
     } else {
-      cuda = std::make_unique<cu::CudaEvent>();
+      cuda = std::make_unique<cu::CopyableCudaEvent>();
     }
   }
 };
@@ -225,7 +264,7 @@ void Event::wait() {
     assert(value() == 1);
     event->cuda->wait();
   } else {
-    event->shared->wait(value());
+    event->atomic->wait(value());
   }
 }
 
@@ -236,7 +275,7 @@ void Event::wait(Stream s) {
     assert(value() == 1);
     event->cuda->wait(s);
   } else {
-    event->shared->wait(s, value());
+    event->atomic->wait(s, value());
   }
 }
 
@@ -247,7 +286,7 @@ void Event::signal(Stream s) {
     assert(value() == 1);
     event->cuda->record(s);
   } else {
-    event->shared->signal(s, value());
+    event->atomic->signal(s, value());
   }
 }
 
@@ -258,9 +297,9 @@ bool Event::is_signaled() const {
   }
   if (event->cuda) {
     assert(value() == 1);
-    return event->cuda->recorded() && event->cuda->completed();
+    return event->cuda->is_signaled();
   } else {
-    return event->shared->is_signaled(value());
+    return event->atomic->is_signaled(value());
   }
 }
 

mlx/backend/cuda/event.h

@@ -3,49 +3,54 @@
 #pragma once
 
 #include "mlx/allocator.h"
+#include "mlx/backend/cuda/utils.h"
 #include "mlx/stream.h"
 
+#include <memory>
+
 #include <cuda_runtime.h>
 #include <cuda/atomic>
 
-#include <memory>
-
 namespace mlx::core::cu {
 
-class CudaEventHandle;
+// RAII-managed move-only wrapper of cudaEvent_t.
+struct CudaEventHandle : public CudaHandle<cudaEvent_t, cudaEventDestroy> {
+  CudaEventHandle(int flags);
+  int flags;
+};
 
 // Wrapper of native cuda event. It can synchronize between GPU streams, or wait
 // on GPU stream in CPU stream, but can not wait on CPU stream.
 class CudaEvent {
  public:
-  CudaEvent();
+  explicit CudaEvent(int flags);
+  ~CudaEvent();
+
+  CudaEvent(CudaEvent&&) = default;
+  CudaEvent& operator=(CudaEvent&&) = default;
+
+  CudaEvent(const CudaEvent&) = delete;
+  CudaEvent& operator=(const CudaEvent&) = delete;
 
   void wait();
   void wait(cudaStream_t stream);
-  void wait(Stream s);
   void record(cudaStream_t stream);
-  void record(Stream s);
 
   // Return whether the recorded kernels have completed. Note that this method
   // returns true if record() has not been called.
   bool completed() const;
 
-  bool recorded() const {
-    return recorded_;
-  }
-
  private:
-  bool recorded_{false};
+  CudaEventHandle event_;
-  std::shared_ptr<CudaEventHandle> event_;
 };
 
 // Event that can synchronize between CPU and GPU. It is much slower than
 // CudaEvent so the latter should always be preferred when possible.
-class SharedEvent {
+class AtomicEvent {
  public:
   using Atomic = cuda::atomic<uint64_t>;
 
-  SharedEvent();
+  AtomicEvent();
 
   void wait(uint64_t value);
   void wait(cudaStream_t stream, uint64_t value);
@@ -57,7 +62,11 @@ class SharedEvent {
   uint64_t value() const;
 
  private:
-  std::shared_ptr<mlx::core::allocator::Buffer> buf_;
+  Atomic* atomic() const {
+    return static_cast<AtomicEvent::Atomic*>(buf_->raw_ptr());
+  }
+
+  std::shared_ptr<allocator::Buffer> buf_;
 };
 
 } // namespace mlx::core::cu

mlx/backend/cuda/fence.cpp

@@ -7,7 +7,7 @@ namespace mlx::core {
 
 struct FenceImpl {
   uint32_t count;
-  cu::SharedEvent event;
+  cu::AtomicEvent event;
 };
 
 Fence::Fence(Stream s) {

mlx/backend/cuda/gemms/cublas_batched_gemm_12_9.cu

@@ -1,208 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/gemms/cublas_gemm.h"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-
-#include <cooperative_groups.h>
-
-namespace mlx::core::cu {
-
-namespace cg = cooperative_groups;
-
-__global__ void set_mm_device_pointers(
-    int8_t** pointers,
-    int8_t* a_start,
-    int8_t* b_start,
-    int8_t* out_start,
-    int item_size,
-    const __grid_constant__ Shape batch_shape,
-    const __grid_constant__ Strides a_batch_strides,
-    const __grid_constant__ Strides b_batch_strides,
-    int64_t batch_stride,
-    int batch_ndim,
-    int batch_count) {
-  auto index = cg::this_grid().thread_rank();
-  if (index >= batch_count) {
-    return;
-  }
-  auto [a_offset, b_offset] = elem_to_loc(
-      index,
-      batch_shape.data(),
-      a_batch_strides.data(),
-      b_batch_strides.data(),
-      batch_ndim);
-  pointers[index] = a_start + item_size * a_offset;
-  pointers[index + batch_count] = b_start + item_size * b_offset;
-  pointers[index + 2 * batch_count] =
-      out_start + item_size * index * batch_stride;
-}
-
-__global__ void set_addmm_device_pointers(
-    int8_t** pointers,
-    int8_t* a_start,
-    int8_t* b_start,
-    int8_t* c_start,
-    int8_t* out_start,
-    int item_size,
-    const __grid_constant__ Shape batch_shape,
-    const __grid_constant__ Strides a_batch_strides,
-    const __grid_constant__ Strides b_batch_strides,
-    const __grid_constant__ Strides c_batch_strides,
-    int64_t batch_stride,
-    int batch_ndim,
-    int batch_count) {
-  auto index = cg::this_grid().thread_rank();
-  if (index >= batch_count) {
-    return;
-  }
-  auto [a_offset, b_offset, c_offset] = elem_to_loc(
-      index,
-      batch_shape.data(),
-      a_batch_strides.data(),
-      b_batch_strides.data(),
-      c_batch_strides.data(),
-      batch_ndim);
-  pointers[index] = a_start + item_size * a_offset;
-  pointers[index + batch_count] = b_start + item_size * b_offset;
-  pointers[index + 2 * batch_count] = c_start + item_size * c_offset;
-  pointers[index + 3 * batch_count] =
-      out_start + item_size * index * batch_stride;
-}
-
-void set_pointer_mode(cublasLtMatrixLayout_t desc, int batch_count) {
-  auto batch_mode = CUBLASLT_BATCH_MODE_POINTER_ARRAY;
-  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
-      desc,
-      CUBLASLT_MATRIX_LAYOUT_BATCH_MODE,
-      &batch_mode,
-      sizeof(batch_mode)));
-  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
-      desc, CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT, &batch_count, sizeof(int32_t)));
-}
-
-void Matmul::run_batched(
-    cu::CommandEncoder& encoder,
-    array& out,
-    const array& a,
-    const array& b,
-    const mlx::core::Shape& batch_shape,
-    const mlx::core::Strides& a_batch_strides,
-    const mlx::core::Strides& b_batch_strides) {
-  auto batch_count = out.size() / (M_ * N_);
-  set_pointer_mode(a_desc_, batch_count);
-  set_pointer_mode(b_desc_, batch_count);
-  set_pointer_mode(out_desc_, batch_count);
-
-  // Launch kernel to set device offsets
-  auto pointers = array(
-      allocator::malloc(batch_count * sizeof(uint64_t) * 3),
-      {static_cast<int>(batch_count * 3)},
-      uint64);
-
-  encoder.add_temporary(pointers);
-  int block_size = 512;
-  encoder.set_output_array(pointers);
-
-  encoder.add_kernel_node(
-      cu::set_mm_device_pointers,
-      cuda::ceil_div(pointers.size(), block_size),
-      block_size,
-      0,
-      pointers.data<int8_t*>(),
-      a.data<int8_t>(),
-      b.data<int8_t>(),
-      out.data<int8_t>(),
-      static_cast<int>(out.dtype().size()),
-      const_param(batch_shape),
-      const_param(a_batch_strides),
-      const_param(b_batch_strides),
-      static_cast<int64_t>(M_) * N_,
-      static_cast<int>(batch_shape.size()),
-      batch_count);
-
-  // Run matmul
-  encoder.set_input_array(pointers);
-  encoder.set_input_array(a);
-  encoder.set_input_array(b);
-  encoder.set_output_array(out);
-
-  auto a_pointers = pointers.data<int8_t*>();
-  auto b_pointers = a_pointers + batch_count;
-  auto out_pointers = b_pointers + batch_count;
-  run_impl(
-      encoder,
-      reinterpret_cast<void*>(out_pointers),
-      reinterpret_cast<void*>(a_pointers),
-      reinterpret_cast<void*>(b_pointers),
-      nullptr);
-}
-
-void Matmul::run_batched(
-    cu::CommandEncoder& encoder,
-    array& out,
-    const array& a,
-    const array& b,
-    const array& c,
-    const mlx::core::Shape& batch_shape,
-    const mlx::core::Strides& a_batch_strides,
-    const mlx::core::Strides& b_batch_strides,
-    const mlx::core::Strides& c_batch_strides,
-    float alpha,
-    float beta) {
-  auto batch_count = out.size() / (M_ * N_);
-  set_pointer_mode(a_desc_, batch_count);
-  set_pointer_mode(b_desc_, batch_count);
-  set_pointer_mode(c_desc_, batch_count);
-  set_pointer_mode(out_desc_, batch_count);
-
-  // Launch kernel to set device offsets
-  auto pointers = array(
-      allocator::malloc(batch_count * sizeof(uint64_t) * 4),
-      {static_cast<int>(batch_count * 4)},
-      uint64);
-
-  encoder.add_temporary(pointers);
-  int block_size = 512;
-  encoder.set_output_array(pointers);
-  encoder.add_kernel_node(
-      cu::set_addmm_device_pointers,
-      cuda::ceil_div(pointers.size(), block_size),
-      block_size,
-      0,
-      pointers.data<int8_t*>(),
-      a.data<int8_t>(),
-      b.data<int8_t>(),
-      c.data<int8_t>(),
-      out.data<int8_t>(),
-      static_cast<int>(out.dtype().size()),
-      const_param(batch_shape),
-      const_param(a_batch_strides),
-      const_param(b_batch_strides),
-      const_param(c_batch_strides),
-      static_cast<int64_t>(M_) * N_,
-      static_cast<int>(batch_shape.size()),
-      batch_count);
-
-  // Run matmul
-  encoder.set_input_array(pointers);
-  encoder.set_input_array(a);
-  encoder.set_input_array(b);
-  encoder.set_input_array(c);
-  encoder.set_output_array(out);
-
-  auto a_pointers = pointers.data<int8_t*>();
-  auto b_pointers = a_pointers + batch_count;
-  auto c_pointers = b_pointers + batch_count;
-  auto out_pointers = c_pointers + batch_count;
-  run_impl(
-      encoder,
-      reinterpret_cast<void*>(out_pointers),
-      reinterpret_cast<void*>(a_pointers),
-      reinterpret_cast<void*>(b_pointers),
-      reinterpret_cast<void*>(c_pointers),
-      alpha,
-      beta);
-}
-
-} // namespace mlx::core::cu

mlx/backend/cuda/gemms/cublas_gemm.cpp

@@ -7,10 +7,12 @@
 
 #include <fmt/format.h>
 
-namespace mlx::core::cu {
+namespace mlx::core {
+
+namespace {
 
 struct CublasPreference {
-  CublasPreference(Device& device) {
+  CublasPreference(cu::Device& device) {
     // The recommended cublas workspace size is 4 MiB for pre-Hopper and 32 MiB
     // for Hopper+:
     // https://docs.nvidia.com/cuda/cublas/#cublassetworkspace
@@ -33,7 +35,7 @@ struct CublasPreference {
   cublasLtMatmulPreference_t pref_{nullptr};
 };
 
-cublasLtMatmulPreference_t cublas_preference(Device& device) {
+cublasLtMatmulPreference_t cublas_preference(cu::Device& device) {
   static CublasPreference pref(device);
   return pref.pref_;
 }
@@ -52,7 +54,7 @@ cublasComputeType_t dtype_to_compute_type(Dtype dtype) {
       return CUBLAS_COMPUTE_64F;
     default:
       throw std::runtime_error(fmt::format(
-          "Unsupported dtype in Matmul: {}.", dtype_to_string(dtype)));
+          "Unsupported dtype in CublasGemm: {}.", dtype_to_string(dtype)));
   }
 }
 
@@ -70,7 +72,7 @@ cudaDataType_t dtype_to_cublas_type(Dtype dtype) {
       return CUDA_C_32F;
     default:
       throw std::runtime_error(fmt::format(
-          "Unsupported dtype in Matmul: {}.", dtype_to_string(dtype)));
+          "Unsupported dtype in CublasGemm: {}.", dtype_to_string(dtype)));
   }
 }
 
@@ -83,10 +85,10 @@ cublasLtMatrixLayout_t create_matrix_layout(
     int32_t batch_count,
     int64_t batch_stride) {
   cublasLtMatrixLayout_t desc;
+  if (transposed) {
+    std::swap(rows, cols);
+  }
   CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutCreate(&desc, type, rows, cols, ld));
-  cublasLtOrder_t order = transposed ? CUBLASLT_ORDER_COL : CUBLASLT_ORDER_ROW;
-  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
-      desc, CUBLASLT_MATRIX_LAYOUT_ORDER, &order, sizeof(cublasLtOrder_t)));
   if (batch_count > 1) {
     CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
         desc,
@@ -102,8 +104,10 @@ cublasLtMatrixLayout_t create_matrix_layout(
   return desc;
 }
 
-Matmul::Matmul(
+} // namespace
-    Device& device,
+
+CublasGemm::CublasGemm(
+    cu::Device& device,
     Dtype dtype,
     bool a_transposed,
     uint64_t a_rows,
@@ -134,29 +138,38 @@ Matmul::Matmul(
       CUBLASLT_MATMUL_DESC_POINTER_MODE,
       &pointer_mode,
       sizeof(int32_t)));
-  cublasOperation_t op = CUBLAS_OP_N;
+
+  // In cublasLt matrices use column-major layout, while it is possible to use
+  // the CUBLASLT_ORDER_ROW option to switch to row-major layout, the bias
+  // epilogue does not work with the option. So instead we swap A and B to make
+  // cublasLt return the row-major result, which works because:
+  // - the data of a matrix in row-major layout is identical to its transpose in
+  //   column-major layout
+  // - C^T = (A @ B)^T = B^T @ A^T
+  cublasOperation_t a_op = b_transposed ? CUBLAS_OP_T : CUBLAS_OP_N;
   CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
       matmul_desc_,
       CUBLASLT_MATMUL_DESC_TRANSA,
-      &op,
+      &a_op,
       sizeof(cublasOperation_t)));
+  cublasOperation_t b_op = a_transposed ? CUBLAS_OP_T : CUBLAS_OP_N;
   CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
       matmul_desc_,
       CUBLASLT_MATMUL_DESC_TRANSB,
-      &op,
+      &b_op,
       sizeof(cublasOperation_t)));
 
   auto type = dtype_to_cublas_type(dtype);
   a_desc_ = create_matrix_layout(
-      type, a_rows, a_cols, a_transposed, lda, batch_count, a_batch_stride);
+      type, b_cols, b_rows, b_transposed, ldb, batch_count, b_batch_stride);
   b_desc_ = create_matrix_layout(
-      type, b_rows, b_cols, b_transposed, ldb, batch_count, b_batch_stride);
+      type, a_cols, a_rows, a_transposed, lda, batch_count, a_batch_stride);
   out_desc_ = create_matrix_layout(
-      type, a_rows, b_cols, false, b_cols, batch_count, a_rows * b_cols);
+      type, b_cols, a_rows, false, b_cols, batch_count, a_rows * b_cols);
 }
 
-Matmul::Matmul(
+CublasGemm::CublasGemm(
-    Device& device,
+    cu::Device& device,
     Dtype dtype,
     bool a_transposed,
     uint64_t a_rows,
@@ -171,7 +184,7 @@ Matmul::Matmul(
     int64_t a_batch_stride,
     int64_t b_batch_stride,
     int64_t c_batch_stride)
-    : Matmul(
+    : CublasGemm(
           device,
           dtype,
           a_transposed,
@@ -187,10 +200,10 @@ Matmul::Matmul(
           b_batch_stride) {
   auto type = dtype_to_cublas_type(dtype);
   c_desc_ = create_matrix_layout(
-      type, a_rows, b_cols, false, ldc, batch_count, c_batch_stride);
+      type, b_cols, a_rows, false, ldc, batch_count, c_batch_stride);
 }
 
-Matmul::~Matmul() {
+CublasGemm::~CublasGemm() {
   CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(a_desc_));
   CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(b_desc_));
   CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(c_desc_));
@@ -198,7 +211,122 @@ Matmul::~Matmul() {
   CHECK_CUBLAS_ERROR(cublasLtMatmulDescDestroy(matmul_desc_));
 }
 
-void Matmul::run_impl(
+void CublasGemm::set_out(
+    Dtype dtype,
+    bool transposed,
+    uint64_t rows,
+    uint64_t cols,
+    int64_t ld,
+    int32_t batch_count,
+    int64_t batch_stride) {
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(out_desc_));
+  out_desc_ = create_matrix_layout(
+      dtype_to_cublas_type(dtype),
+      cols,
+      rows,
+      transposed,
+      ld,
+      batch_count,
+      batch_stride);
+}
+
+void CublasGemm::set_bias(cu::CommandEncoder& encoder, const array& bias) {
+  encoder.set_input_array(bias);
+  cublasLtEpilogue_t epilogue = CUBLASLT_EPILOGUE_BIAS;
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+      matmul_desc_,
+      CUBLASLT_MATMUL_DESC_EPILOGUE,
+      &epilogue,
+      sizeof(epilogue)));
+  auto* bias_ptr = bias.data<void>();
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+      matmul_desc_,
+      CUBLASLT_MATMUL_DESC_BIAS_POINTER,
+      &bias_ptr,
+      sizeof(bias_ptr)));
+}
+
+void CublasGemm::run(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const Shape& batch_shape,
+    const Strides& a_batch_strides,
+    const Strides& b_batch_strides,
+    float alpha) {
+  int batch_count = out.size() / (M_ * N_);
+  if (batch_count / batch_shape.back() > 1) {
+    run_batched(
+        encoder,
+        out,
+        a,
+        b,
+        batch_shape,
+        a_batch_strides,
+        b_batch_strides,
+        alpha);
+    return;
+  }
+
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+
+  execute(
+      encoder,
+      out.data<void>(),
+      a.data<void>(),
+      b.data<void>(),
+      nullptr,
+      alpha);
+}
+
+void CublasGemm::run(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const array& c,
+    const Shape& batch_shape,
+    const Strides& a_batch_strides,
+    const Strides& b_batch_strides,
+    const Strides& c_batch_strides,
+    float alpha,
+    float beta) {
+  int batch_count = out.size() / (M_ * N_);
+  if (batch_count / batch_shape.back() > 1) {
+    run_batched(
+        encoder,
+        out,
+        a,
+        b,
+        c,
+        batch_shape,
+        a_batch_strides,
+        b_batch_strides,
+        c_batch_strides,
+        alpha,
+        beta);
+    return;
+  }
+
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(c);
+  encoder.set_output_array(out);
+
+  execute(
+      encoder,
+      out.data<void>(),
+      a.data<void>(),
+      b.data<void>(),
+      c.data<void>(),
+      alpha,
+      beta);
+}
+
+void CublasGemm::execute(
     cu::CommandEncoder& encoder,
     void* out,
     const void* a,
@@ -241,9 +369,9 @@ void Matmul::run_impl(
       handle_,
       matmul_desc_,
       &alpha,
-      a,
+      b, // a and b are swapped
       a_desc_,
-      b,
+      a,
       b_desc_,
       &beta,
       c ? c : out,
@@ -256,29 +384,4 @@ void Matmul::run_impl(
       encoder.stream()));
 }
 
-void Matmul::run(
+} // namespace mlx::core
-    cu::CommandEncoder& encoder,
-    array& out,
-    const array& a,
-    const array& b,
-    const std::optional<array>& c /* = std::nullopt */,
-    float alpha /* = 1 */,
-    float beta /* = 0 */) {
-  encoder.set_input_array(a);
-  encoder.set_input_array(b);
-  if (c) {
-    encoder.set_input_array(*c);
-  }
-  encoder.set_output_array(out);
-
-  run_impl(
-      encoder,
-      out.data<void>(),
-      a.data<void>(),
-      b.data<void>(),
-      c ? c->data<void>() : nullptr,
-      alpha,
-      beta);
-}
-
-} // namespace mlx::core::cu

mlx/backend/cuda/gemms/cublas_gemm.h

@@ -5,13 +5,13 @@
 #include "mlx/backend/cuda/device.h"
 
 #include <cublasLt.h>
-#include <optional>
 
-namespace mlx::core::cu {
+namespace mlx::core {
-class Matmul {
+
+class CublasGemm {
  public:
-  Matmul(
+  CublasGemm(
-      Device& device,
+      cu::Device& device,
       Dtype dtype,
       bool a_transposed,
       uint64_t a_rows,
@@ -25,8 +25,8 @@ class Matmul {
       int64_t a_batch_stride,
       int64_t b_batch_stride);
 
-  Matmul(
+  CublasGemm(
-      Device& device,
+      cu::Device& device,
       Dtype dtype,
       bool a_transposed,
       uint64_t a_rows,
@@ -42,25 +42,54 @@ class Matmul {
       int64_t b_batch_stride,
       int64_t c_batch_stride);
 
-  ~Matmul();
+  ~CublasGemm();
+
+  // The output's descriptor is inferred from inputs by default, use this method
+  // for unusual output.
+  void set_out(
+      Dtype dtype,
+      bool transposed,
+      uint64_t rows,
+      uint64_t cols,
+      int64_t ld,
+      int32_t batch_count,
+      int64_t batch_stride);
+
+  void set_bias(cu::CommandEncoder& encoder, const array& bias);
 
   void run(
       cu::CommandEncoder& encoder,
       array& out,
       const array& a,
       const array& b,
-      const std::optional<array>& c = std::nullopt,
+      const Shape& batch_shape,
-      float alpha = 1,
+      const Strides& a_batch_strides,
-      float beta = 0);
+      const Strides& b_batch_strides,
+      float alpha = 1.0f);
 
+  void run(
+      cu::CommandEncoder& encoder,
+      array& out,
+      const array& a,
+      const array& b,
+      const array& c,
+      const Shape& batch_shape,
+      const Strides& a_batch_strides,
+      const Strides& b_batch_strides,
+      const Strides& c_batch_strides,
+      float alpha,
+      float beta);
+
+ private:
   void run_batched(
       cu::CommandEncoder& encoder,
       array& out,
       const array& a,
       const array& b,
-      const mlx::core::Shape& batch_shape,
+      const Shape& batch_shape,
-      const mlx::core::Strides& a_batch_strides,
+      const Strides& a_batch_strides,
-      const mlx::core::Strides& b_batch_strides);
+      const Strides& b_batch_strides,
+      float alpha);
 
   void run_batched(
       cu::CommandEncoder& encoder,
@@ -68,15 +97,14 @@ class Matmul {
       const array& a,
       const array& b,
       const array& c,
-      const mlx::core::Shape& batch_shape,
+      const Shape& batch_shape,
-      const mlx::core::Strides& a_batch_strides,
+      const Strides& a_batch_strides,
-      const mlx::core::Strides& b_batch_strides,
+      const Strides& b_batch_strides,
-      const mlx::core::Strides& c_batch_strides,
+      const Strides& c_batch_strides,
       float alpha,
       float beta);
 
- private:
+  void execute(
-  void run_impl(
       cu::CommandEncoder& encoder,
       void* out,
       const void* a,
@@ -97,4 +125,4 @@ class Matmul {
   cublasLtMatmulHeuristicResult_t heuristic_;
 };
 
-} // namespace mlx::core::cu
+} // namespace mlx::core

mlx/backend/cuda/gemms/cublas_gemm_batched_12_0.cpp

@@ -4,16 +4,17 @@
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/gemms/cublas_gemm.h"
 
-namespace mlx::core::cu {
+namespace mlx::core {
 
-void Matmul::run_batched(
+void CublasGemm::run_batched(
     cu::CommandEncoder& encoder,
     array& out,
     const array& a,
     const array& b,
-    const mlx::core::Shape& batch_shape,
+    const Shape& batch_shape,
-    const mlx::core::Strides& a_batch_strides,
+    const Strides& a_batch_strides,
-    const mlx::core::Strides& b_batch_strides) {
+    const Strides& b_batch_strides,
+    float alpha) {
   encoder.set_input_array(a);
   encoder.set_input_array(b);
   encoder.set_output_array(out);
@@ -22,27 +23,28 @@ void Matmul::run_batched(
   ContiguousIterator b_it(batch_shape, b_batch_strides, batch_shape.size() - 1);
   auto concurrent = encoder.concurrent_context();
   for (size_t i = 0; i < nbatch; ++i) {
-    run_impl(
+    execute(
         encoder,
         out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M_ * N_,
         a.data<int8_t>() + a.itemsize() * a_it.loc,
         b.data<int8_t>() + b.itemsize() * b_it.loc,
-        nullptr);
+        nullptr,
+        alpha);
     a_it.step();
     b_it.step();
   }
 }
 
-void Matmul::run_batched(
+void CublasGemm::run_batched(
     cu::CommandEncoder& encoder,
     array& out,
     const array& a,
     const array& b,
     const array& c,
-    const mlx::core::Shape& batch_shape,
+    const Shape& batch_shape,
-    const mlx::core::Strides& a_batch_strides,
+    const Strides& a_batch_strides,
-    const mlx::core::Strides& b_batch_strides,
+    const Strides& b_batch_strides,
-    const mlx::core::Strides& c_batch_strides,
+    const Strides& c_batch_strides,
     float alpha,
     float beta) {
   encoder.set_input_array(a);
@@ -56,7 +58,7 @@ void Matmul::run_batched(
   ContiguousIterator c_it(batch_shape, c_batch_strides, batch_shape.size() - 1);
   auto concurrent = encoder.concurrent_context();
   for (size_t i = 0; i < nbatch; ++i) {
-    run_impl(
+    execute(
         encoder,
         out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M_ * N_,
         a.data<int8_t>() + a.itemsize() * a_it.loc,
@@ -70,4 +72,4 @@ void Matmul::run_batched(
   }
 }
 
-} // namespace mlx::core::cu
+} // namespace mlx::core

mlx/backend/cuda/gemms/cublas_gemm_batched_12_9.cu

@@ -0,0 +1,329 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/gemms/cublas_gemm.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <int NDIM>
+__global__ void set_mm_device_pointers_nd(
+    int8_t** pointers,
+    int8_t* a_start,
+    int8_t* b_start,
+    int8_t* out_start,
+    int item_size,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> batch_shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> a_batch_strides,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> b_batch_strides,
+    int64_t batch_stride,
+    int batch_count) {
+  auto index = cg::this_grid().thread_rank();
+  if (index >= batch_count) {
+    return;
+  }
+  auto [a_offset, b_offset] = elem_to_loc_nd<NDIM>(
+      index,
+      batch_shape.data(),
+      a_batch_strides.data(),
+      b_batch_strides.data());
+  pointers[index] = a_start + item_size * a_offset;
+  pointers[index + batch_count] = b_start + item_size * b_offset;
+  pointers[index + 2 * batch_count] =
+      out_start + item_size * index * batch_stride;
+}
+
+__global__ void set_mm_device_pointers_g(
+    int8_t** pointers,
+    int8_t* a_start,
+    int8_t* b_start,
+    int8_t* out_start,
+    int item_size,
+    const __grid_constant__ Shape batch_shape,
+    const __grid_constant__ Strides a_batch_strides,
+    const __grid_constant__ Strides b_batch_strides,
+    int64_t batch_stride,
+    int batch_ndim,
+    int batch_count) {
+  auto index = cg::this_grid().thread_rank();
+  if (index >= batch_count) {
+    return;
+  }
+  auto [a_offset, b_offset] = elem_to_loc(
+      index,
+      batch_shape.data(),
+      a_batch_strides.data(),
+      b_batch_strides.data(),
+      batch_ndim);
+  pointers[index] = a_start + item_size * a_offset;
+  pointers[index + batch_count] = b_start + item_size * b_offset;
+  pointers[index + 2 * batch_count] =
+      out_start + item_size * index * batch_stride;
+}
+
+template <int NDIM>
+__global__ void set_addmm_device_pointers_nd(
+    int8_t** pointers,
+    int8_t* a_start,
+    int8_t* b_start,
+    int8_t* c_start,
+    int8_t* out_start,
+    int item_size,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> batch_shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> a_batch_strides,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> b_batch_strides,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> c_batch_strides,
+    int64_t batch_stride,
+    int batch_count) {
+  auto index = cg::this_grid().thread_rank();
+  if (index >= batch_count) {
+    return;
+  }
+  auto [a_offset, b_offset, c_offset] = elem_to_loc_nd<NDIM>(
+      index,
+      batch_shape.data(),
+      a_batch_strides.data(),
+      b_batch_strides.data(),
+      c_batch_strides.data());
+  pointers[index] = a_start + item_size * a_offset;
+  pointers[index + batch_count] = b_start + item_size * b_offset;
+  pointers[index + 2 * batch_count] = c_start + item_size * c_offset;
+  pointers[index + 3 * batch_count] =
+      out_start + item_size * index * batch_stride;
+}
+
+__global__ void set_addmm_device_pointers_g(
+    int8_t** pointers,
+    int8_t* a_start,
+    int8_t* b_start,
+    int8_t* c_start,
+    int8_t* out_start,
+    int item_size,
+    const __grid_constant__ Shape batch_shape,
+    const __grid_constant__ Strides a_batch_strides,
+    const __grid_constant__ Strides b_batch_strides,
+    const __grid_constant__ Strides c_batch_strides,
+    int64_t batch_stride,
+    int batch_ndim,
+    int batch_count) {
+  auto index = cg::this_grid().thread_rank();
+  if (index >= batch_count) {
+    return;
+  }
+  auto [a_offset, b_offset, c_offset] = elem_to_loc(
+      index,
+      batch_shape.data(),
+      a_batch_strides.data(),
+      b_batch_strides.data(),
+      c_batch_strides.data(),
+      batch_ndim);
+  pointers[index] = a_start + item_size * a_offset;
+  pointers[index + batch_count] = b_start + item_size * b_offset;
+  pointers[index + 2 * batch_count] = c_start + item_size * c_offset;
+  pointers[index + 3 * batch_count] =
+      out_start + item_size * index * batch_stride;
+}
+
+} // namespace cu
+
+namespace {
+
+void set_pointer_mode(cublasLtMatrixLayout_t desc, int batch_count) {
+  auto batch_mode = CUBLASLT_BATCH_MODE_POINTER_ARRAY;
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+      desc,
+      CUBLASLT_MATRIX_LAYOUT_BATCH_MODE,
+      &batch_mode,
+      sizeof(batch_mode)));
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+      desc, CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT, &batch_count, sizeof(int32_t)));
+}
+
+} // namespace
+
+void CublasGemm::run_batched(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const Shape& batch_shape,
+    const Strides& a_batch_strides,
+    const Strides& b_batch_strides,
+    float alpha) {
+  int batch_count = out.size() / (M_ * N_);
+  set_pointer_mode(a_desc_, batch_count);
+  set_pointer_mode(b_desc_, batch_count);
+  set_pointer_mode(out_desc_, batch_count);
+
+  // Launch kernel to set device offsets
+  auto pointers = array(
+      allocator::malloc(batch_count * sizeof(void*) * 3),
+      {batch_count * 3},
+      uint64);
+
+  encoder.add_temporary(pointers);
+  encoder.set_output_array(pointers);
+
+  int block_dims = std::min(batch_count, 256);
+  int num_blocks = cuda::ceil_div(batch_count, block_dims);
+  int64_t batch_stride = M_ * N_;
+  int item_size = out.itemsize();
+
+  int ndim = batch_shape.size();
+  if (ndim <= 3) {
+    dispatch_1_2_3(ndim, [&](auto ndim_constant) {
+      encoder.add_kernel_node(
+          cu::set_mm_device_pointers_nd<ndim_constant()>,
+          num_blocks,
+          block_dims,
+          0,
+          pointers.data<int8_t*>(),
+          a.data<int8_t>(),
+          b.data<int8_t>(),
+          out.data<int8_t>(),
+          item_size,
+          const_param<ndim_constant()>(batch_shape),
+          const_param<ndim_constant()>(a_batch_strides),
+          const_param<ndim_constant()>(b_batch_strides),
+          batch_stride,
+          batch_count);
+    });
+  } else {
+    encoder.add_kernel_node(
+        cu::set_mm_device_pointers_g,
+        num_blocks,
+        block_dims,
+        0,
+        pointers.data<int8_t*>(),
+        a.data<int8_t>(),
+        b.data<int8_t>(),
+        out.data<int8_t>(),
+        item_size,
+        const_param(batch_shape),
+        const_param(a_batch_strides),
+        const_param(b_batch_strides),
+        batch_stride,
+        ndim,
+        batch_count);
+  }
+
+  // Run matmul
+  encoder.set_input_array(pointers);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+
+  auto a_pointers = pointers.data<int8_t*>();
+  auto b_pointers = a_pointers + batch_count;
+  auto out_pointers = b_pointers + batch_count;
+  execute(
+      encoder,
+      reinterpret_cast<void*>(out_pointers),
+      reinterpret_cast<void*>(a_pointers),
+      reinterpret_cast<void*>(b_pointers),
+      nullptr,
+      alpha);
+}
+
+void CublasGemm::run_batched(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const array& c,
+    const Shape& batch_shape,
+    const Strides& a_batch_strides,
+    const Strides& b_batch_strides,
+    const Strides& c_batch_strides,
+    float alpha,
+    float beta) {
+  int batch_count = out.size() / (M_ * N_);
+  set_pointer_mode(a_desc_, batch_count);
+  set_pointer_mode(b_desc_, batch_count);
+  set_pointer_mode(c_desc_, batch_count);
+  set_pointer_mode(out_desc_, batch_count);
+
+  // Launch kernel to set device offsets
+  auto pointers = array(
+      allocator::malloc(batch_count * sizeof(uint64_t) * 4),
+      {batch_count * 4},
+      uint64);
+
+  encoder.add_temporary(pointers);
+  encoder.set_output_array(pointers);
+
+  int block_dims = std::min(batch_count, 256);
+  int num_blocks = cuda::ceil_div(batch_count, block_dims);
+  int64_t batch_stride = M_ * N_;
+  int item_size = out.itemsize();
+
+  int ndim = batch_shape.size();
+  if (ndim <= 3) {
+    dispatch_1_2_3(ndim, [&](auto ndim_constant) {
+      encoder.add_kernel_node(
+          cu::set_addmm_device_pointers_nd<ndim_constant()>,
+          num_blocks,
+          block_dims,
+          0,
+          pointers.data<int8_t*>(),
+          a.data<int8_t>(),
+          b.data<int8_t>(),
+          c.data<int8_t>(),
+          out.data<int8_t>(),
+          item_size,
+          const_param<ndim_constant()>(batch_shape),
+          const_param<ndim_constant()>(a_batch_strides),
+          const_param<ndim_constant()>(b_batch_strides),
+          const_param<ndim_constant()>(c_batch_strides),
+          batch_stride,
+          batch_count);
+    });
+  } else {
+    encoder.add_kernel_node(
+        cu::set_addmm_device_pointers_g,
+        num_blocks,
+        block_dims,
+        0,
+        pointers.data<int8_t*>(),
+        a.data<int8_t>(),
+        b.data<int8_t>(),
+        c.data<int8_t>(),
+        out.data<int8_t>(),
+        item_size,
+        const_param(batch_shape),
+        const_param(a_batch_strides),
+        const_param(b_batch_strides),
+        const_param(c_batch_strides),
+        batch_stride,
+        ndim,
+        batch_count);
+  }
+
+  // Run matmul
+  encoder.set_input_array(pointers);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(c);
+  encoder.set_output_array(out);
+
+  auto a_pointers = pointers.data<int8_t*>();
+  auto b_pointers = a_pointers + batch_count;
+  auto c_pointers = b_pointers + batch_count;
+  auto out_pointers = c_pointers + batch_count;
+  execute(
+      encoder,
+      reinterpret_cast<void*>(out_pointers),
+      reinterpret_cast<void*>(a_pointers),
+      reinterpret_cast<void*>(b_pointers),
+      reinterpret_cast<void*>(c_pointers),
+      alpha,
+      beta);
+}
+
+} // namespace mlx::core

mlx/backend/cuda/indexing.cpp

@@ -94,7 +94,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
             large ? "int64_t" : "int32_t"));
       }
     }
-    return std::make_pair(jit_source_gather, std::move(kernel_names));
+    return std::make_tuple(false, jit_source_gather, std::move(kernel_names));
   });
 
   cu::KernelArgs args;
@@ -110,7 +110,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   args.append<int32_t>(src.ndim());
   args.append_ndim(slice_sizes_);
   args.append(slice_size);
-  args.append(SmallVector<int32_t>(axes_.begin(), axes_.end()));
+  args.append(axes_);
   append_indices_arg(args, inputs, nidx, idx_ndim);
 
   std::string kernel_name = fmt::format(
@@ -189,7 +189,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
             large ? "int64_t" : "int32_t"));
       }
     }
-    return std::make_pair(jit_source_scatter, std::move(kernel_names));
+    return std::make_tuple(false, jit_source_scatter, std::move(kernel_names));
   });
 
   cu::KernelArgs args;
@@ -211,7 +211,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   args.append_ndim(out.shape());
   args.append_ndim(out.strides());
   args.append<int32_t>(out.ndim());
-  args.append(SmallVector<int32_t>(axes_.begin(), axes_.end()));
+  args.append(axes_);
   append_indices_arg(args, inputs, nidx, idx_ndim);
 
   std::string kernel_name = fmt::format(
@@ -268,7 +268,8 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
         }
       }
     }
-    return std::make_pair(jit_source_gather_axis, std::move(kernel_names));
+    return std::make_tuple(
+        false, jit_source_gather_axis, std::move(kernel_names));
   });
 
   size_t idx_size_pre = 1;
@@ -371,7 +372,8 @@ void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
         }
       }
     }
-    return std::make_pair(jit_source_scatter_axis, std::move(kernel_names));
+    return std::make_tuple(
+        false, jit_source_scatter_axis, std::move(kernel_names));
   });
 
   size_t idx_size_pre = 1;

mlx/backend/cuda/jit_module.cpp

@@ -67,10 +67,12 @@ const std::string& cccl_dir() {
       return path.string();
     }
     // Finally check the environment variable.
-    path = std::getenv("MLX_CCCL_DIR");
+    if (const char* env = std::getenv("MLX_CCCL_DIR"); env) {
+      path = env;
       if (!path.empty() && std::filesystem::exists(path)) {
         return path.string();
       }
+    }
     return std::string();
   }();
   return dir;
@@ -101,8 +103,8 @@ const std::filesystem::path& ptx_cache_dir() {
 bool read_cached_ptx(
     const std::filesystem::path& cache_dir,
     const std::string& module_name,
-    std::vector<char>* ptx,
+    std::string& ptx,
-    std::vector<std::pair<std::string, std::string>>* ptx_kernels) {
+    std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
   if (cache_dir.empty()) {
     return false;
   }
@@ -117,15 +119,15 @@ bool read_cached_ptx(
   if (!ptx_file.good()) {
     return false;
   }
-  ptx->resize(ptx_size);
+  ptx.resize(ptx_size);
-  ptx_file.read(ptx->data(), ptx_size);
+  ptx_file.read(ptx.data(), ptx_size);
 
   std::ifstream txt_file(cache_dir / (module_name + ".txt"), std::ios::binary);
   std::string line;
   while (std::getline(txt_file, line)) {
     auto tab = line.find('\t');
     if (tab != std::string::npos) {
-      ptx_kernels->emplace_back(line.substr(0, tab), line.substr(tab + 1));
+      ptx_kernels.emplace_back(line.substr(0, tab), line.substr(tab + 1));
     }
   }
   return true;
@@ -135,7 +137,7 @@ bool read_cached_ptx(
 void write_cached_ptx(
     const std::filesystem::path& cache_dir,
     const std::string& module_name,
-    const std::vector<char>& ptx,
+    const std::string& ptx,
     const std::vector<std::pair<std::string, std::string>>& ptx_kernels,
     const std::string& source_code) {
   if (cache_dir.empty()) {
@@ -217,22 +219,18 @@ constexpr const char* g_headers[] = {
     jit_source_utils,
 };
 
-} // namespace
+void compile(
-
-JitModule::JitModule(
     Device& device,
     const std::string& module_name,
-    const KernelBuilder& builder) {
+    const std::string& source,
-  // Check cache.
+    const std::vector<std::string>& kernel_names,
-  std::vector<char> ptx;
+    std::string& ptx,
-  std::vector<std::pair<std::string, std::string>> ptx_kernels;
+    std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
-  if (!read_cached_ptx(ptx_cache_dir(), module_name, &ptx, &ptx_kernels)) {
+  // Create the program
-    // Create program.
-    auto [source_code, kernel_names] = builder();
   nvrtcProgram prog;
   CHECK_NVRTC_ERROR(nvrtcCreateProgram(
       &prog,
-        source_code.c_str(),
+      source.c_str(),
       (module_name + ".cu").c_str(),
       std::size(g_headers),
       g_headers,
@@ -286,16 +284,20 @@ JitModule::JitModule(
   } else {
     CHECK_NVRTC_ERROR(nvrtcGetPTXSize(prog, &ptx_size));
   }
-    ptx.resize(ptx_size, 0);
+  ptx.resize(ptx_size);
   if (use_sass) {
     CHECK_NVRTC_ERROR(nvrtcGetCUBIN(prog, ptx.data()));
   } else {
     CHECK_NVRTC_ERROR(nvrtcGetPTX(prog, ptx.data()));
   }
-    write_cached_ptx(
+}
-        ptx_cache_dir(), module_name, ptx, ptx_kernels, source_code);
-  }
 
+void load_module(
+    const std::string& module_name,
+    const std::string& ptx,
+    const std::vector<std::pair<std::string, std::string>>& ptx_kernels,
+    CUmodule& module_,
+    std::unordered_map<std::string, std::pair<CUfunction, bool>>& kernels) {
   // Load module.
   char jit_log[4089] = {};
   CUjit_option options[] = {
@@ -312,21 +314,69 @@ JitModule::JitModule(
   for (const auto& [name, mangled] : ptx_kernels) {
     CUfunction kernel;
     CHECK_CUDA_ERROR(cuModuleGetFunction(&kernel, module_, mangled.c_str()));
-    kernels_[name] = kernel;
+    kernels[name] = std::make_pair(kernel, false);
   }
 }
 
+} // namespace
+
+JitModule::JitModule(
+    Device& device,
+    const std::string& module_name,
+    const KernelBuilder& builder,
+    bool use_disk_cache) {
+  // Will hold the actual device executable source code and kernel names
+  std::string ptx;
+  std::vector<std::pair<std::string, std::string>> ptx_kernels;
+
+  // Try to load them from the file cache
+  if (!read_cached_ptx(ptx_cache_dir(), module_name, ptx, ptx_kernels)) {
+    auto [precompiled, source_code, kernel_names] = builder();
+
+    // Get the PTX or cubin
+    if (precompiled) {
+      ptx = std::move(source_code);
+      for (auto& name : kernel_names) {
+        ptx_kernels.emplace_back(name, name);
+      }
+    } else {
+      compile(device, module_name, source_code, kernel_names, ptx, ptx_kernels);
+    }
+
+    // If requested save them in the file cache for the next launch
+    if (use_disk_cache) {
+      write_cached_ptx(
+          ptx_cache_dir(), module_name, ptx, ptx_kernels, source_code);
+    }
+  }
+
+  // Load the module
+  load_module(module_name, ptx, ptx_kernels, module_, kernels_);
+}
+
 JitModule::~JitModule() {
   CHECK_CUDA_ERROR(cuModuleUnload(module_));
 }
 
-CUfunction JitModule::get_kernel(const std::string& kernel_name) {
+CUfunction JitModule::get_kernel(
+    const std::string& kernel_name,
+    std::function<void(CUfunction)> configure_kernel) {
   auto it = kernels_.find(kernel_name);
   if (it == kernels_.end()) {
     throw std::runtime_error(
         fmt::format("There is no kernel named {}.", kernel_name));
   }
-  return it->second;
+
+  // If it is the first time we run this kernel then configure it. Do it only
+  // once!
+  if (!it->second.second) {
+    if (configure_kernel) {
+      configure_kernel(it->second.first);
+    }
+    it->second.second = true;
+  }
+
+  return it->second.first;
 }
 
 std::unordered_map<std::string, JitModule>& get_jit_module_cache() {
@@ -337,11 +387,12 @@ std::unordered_map<std::string, JitModule>& get_jit_module_cache() {
 JitModule& get_jit_module(
     const mlx::core::Device& device,
     const std::string& name,
-    const KernelBuilder& builder) {
+    const KernelBuilder& builder,
+    bool cache) {
   auto& map = get_jit_module_cache();
   auto it = map.find(name);
   if (it == map.end()) {
-    it = map.try_emplace(name, cu::device(device), name, builder).first;
+    it = map.try_emplace(name, cu::device(device), name, builder, cache).first;
   }
   return it->second;
 }

mlx/backend/cuda/jit_module.h

@@ -19,7 +19,8 @@ namespace mlx::core::cu {
 
 class Device;
 
-using KernelBuilderResult = std::pair<
+using KernelBuilderResult = std::tuple<
+    /* precompiled */ bool,
     /* source code */ std::string,
     /* kernel names */ std::vector<std::string>>;
 using KernelBuilder = std::function<KernelBuilderResult()>;
@@ -45,6 +46,11 @@ struct KernelArgs {
     append_ptr(std::get<SmallVector<T>>(storage_.back()).data());
   }
 
+  template <typename T>
+  void append(const std::vector<T>& vec) {
+    append(SmallVector<T>(vec.begin(), vec.end()));
+  }
+
   // Make sure the arg is copied to an array with size of NDIM.
   template <size_t NDIM = MAX_NDIM, typename T>
   void append_ndim(SmallVector<T> vec) {
@@ -63,14 +69,16 @@ struct KernelArgs {
  private:
   std::vector<void*> args_;
 
-  // The cuLaunchKernel API requires passing pointers to arguments so store
+  // The cuGraphAddKernelNode API requires passing pointers to arguments so
-  // temporary values untill kernel is launched.
+  // store temporary values until the node is created.
   using Arg = std::variant<
       std::monostate,
       CUdeviceptr,
+      bool,
       int32_t,
       uint32_t,
       int64_t,
+      float,
       SmallVector<const void*>,
       SmallVector<int32_t>,
       SmallVector<int64_t>>;
@@ -82,16 +90,19 @@ class JitModule {
   JitModule(
       Device& device,
       const std::string& module_name,
-      const KernelBuilder& builder);
+      const KernelBuilder& builder,
+      bool cache);
   ~JitModule();
 
   JitModule(const JitModule&) = delete;
   JitModule& operator=(const JitModule&) = delete;
-  CUfunction get_kernel(const std::string& kernel_name);
+  CUfunction get_kernel(
+      const std::string& kernel_name,
+      std::function<void(CUfunction)> configure_kernel = nullptr);
 
  private:
   CUmodule module_{nullptr};
-  std::unordered_map<std::string, CUfunction> kernels_;
+  std::unordered_map<std::string, std::pair<CUfunction, bool>> kernels_;
 };
 
 std::unordered_map<std::string, JitModule>& get_jit_module_cache();
@@ -99,6 +110,7 @@ std::unordered_map<std::string, JitModule>& get_jit_module_cache();
 JitModule& get_jit_module(
     const mlx::core::Device& device,
     const std::string& name,
-    const KernelBuilder& builder);
+    const KernelBuilder& builder,
+    bool use_disk_cache = true);
 
 } // namespace mlx::core::cu

mlx/backend/cuda/lru_cache.h

@@ -2,11 +2,15 @@
 
 #pragma once
 
+#include "mlx/utils.h"
+
 #include <cstring>
 #include <list>
 #include <unordered_map>
 #include <utility>
 
+#include <fmt/format.h>
+
 namespace mlx::core {
 
 template <
@@ -27,6 +31,14 @@ class LRUCache {
     }
   }
 
+  // Initialize with capacity read from |env_name|.
+  LRUCache(const char* env_name, int default_capacity)
+      : LRUCache(env::get_var(env_name, default_capacity)) {
+    if (env::get_var("MLX_ENABLE_CACHE_THRASHING_CHECK", 1)) {
+      env_name_ = env_name;
+    }
+  }
+
   size_t size() const {
     return map_.size();
   }
@@ -76,6 +88,14 @@ class LRUCache {
       return {it->second, false};
     }
 
+    if (env_name_ && ++cache_misses_ > 2 * capacity_) {
+      throw std::runtime_error(fmt::format(
+          "Cache thrashing is happening, please set the environment variable "
+          "{} to a larger value than {} to fix degraded performance.",
+          env_name_,
+          capacity_));
+    }
+
     vlist_.emplace_front(key, std::forward<U>(value));
     map_[key] = vlist_.begin();
 
@@ -106,6 +126,9 @@ class LRUCache {
     }
   }
 
+  const char* env_name_{nullptr};
+  size_t cache_misses_{0};
+
   list_type vlist_;
   map_type map_;
   size_t capacity_;

mlx/backend/cuda/matmul.cpp

@@ -11,6 +11,7 @@
 #include <numeric>
 
 namespace mlx::core {
+
 namespace {
 
 std::tuple<bool, int64_t, array>
@@ -28,6 +29,76 @@ check_transpose(cu::CommandEncoder& enc, const Stream& s, const array& arr) {
   }
 }
 
+void gemm_and_bias(
+    cu::CommandEncoder& encoder,
+    int M,
+    int N,
+    int K,
+    bool a_transposed,
+    int64_t lda,
+    bool b_transposed,
+    int64_t ldb,
+    array& out,
+    const array& a,
+    const array& b,
+    const std::optional<array>& bias = std::nullopt,
+    float alpha = 1.0f) {
+  // Check and collapse batch dimensions
+  auto [batch_shape, a_batch_strides, b_batch_strides] = collapse_batches(a, b);
+
+  auto batch_count = out.size() / (M * N);
+
+  // Collapse batches into M if needed
+  if (batch_count > 1 && !a_transposed && batch_shape.size() == 1 &&
+      a.strides()[a.ndim() - 2] == K && a_batch_strides.back() == M * K &&
+      b_batch_strides.back() == 0) {
+    M *= batch_shape.back();
+    batch_count = 1;
+
+    a_batch_strides = {0};
+    b_batch_strides = {0};
+    batch_shape = {1};
+  }
+
+  // Use gemmv when possible
+  if (!bias && cu::can_use_gemv(M, N, K, a_transposed, b_transposed)) {
+    cu::gemv(
+        a,
+        b,
+        out,
+        M,
+        N,
+        K,
+        batch_count,
+        batch_shape,
+        a_batch_strides,
+        b_batch_strides,
+        encoder);
+    return;
+  }
+
+  // Invoke cublasLt
+  CublasGemm gemm(
+      encoder.device(),
+      a.dtype(),
+      a_transposed,
+      M,
+      K,
+      lda,
+      b_transposed,
+      K,
+      N,
+      ldb,
+      batch_shape.back(),
+      a_batch_strides.back(),
+      b_batch_strides.back());
+  if (bias) {
+    gemm.set_bias(encoder, *bias);
+  }
+  gemm.run(
+      encoder, out, a, b, batch_shape, a_batch_strides, b_batch_strides, alpha);
+}
+
 } // namespace
 
 void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -48,9 +119,6 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   out.set_data(allocator::malloc(out.nbytes()));
 
-  /////////////////////////////////////////////////////////////////////////////
-  // Init checks and prep
-
   int M = a_pre.shape(-2);
   int N = b_pre.shape(-1);
   int K = a_pre.shape(-1);
@@ -60,65 +128,8 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto [a_transposed, lda, a] = check_transpose(encoder, s, a_pre);
   auto [b_transposed, ldb, b] = check_transpose(encoder, s, b_pre);
 
-  /////////////////////////////////////////////////////////////////////////////
+  gemm_and_bias(
-  // Check and collapse batch dimensions
+      encoder, M, N, K, a_transposed, lda, b_transposed, ldb, out, a, b);
-
-  auto [batch_shape, a_batch_strides, b_batch_strides] = collapse_batches(a, b);
-
-  auto batch_count = out.size() / (M * N);
-
-  // Collapse batches into M if needed
-  if (batch_count > 1 && !a_transposed && batch_shape.size() == 1 &&
-      a.strides()[a.ndim() - 2] == K && a_batch_strides.back() == M * K &&
-      b_batch_strides.back() == 0) {
-    M *= batch_shape.back();
-    batch_count = 1;
-
-    a_batch_strides = {0};
-    b_batch_strides = {0};
-    batch_shape = {1};
-  }
-
-  if (cu::can_use_gemv(M, N, K, a_transposed, b_transposed)) {
-    cu::gemv(
-        a,
-        b,
-        out,
-        M,
-        N,
-        K,
-        batch_count,
-        batch_shape,
-        a_batch_strides,
-        b_batch_strides,
-        encoder);
-    return;
-  }
-
-  /////////////////////////////////////////////////////////////////////////////
-  // Invoke cublasLt
-  cu::Matmul matmul(
-      cu::device(s.device),
-      a.dtype(),
-      a_transposed,
-      M,
-      K,
-      lda,
-      b_transposed,
-      K,
-      N,
-      ldb,
-      batch_shape.back(),
-      a_batch_strides.back(),
-      b_batch_strides.back());
-
-  if ((batch_count / batch_shape.back()) == 1) {
-    matmul.run(encoder, out, a, b);
-    return;
-  }
-
-  matmul.run_batched(
-      encoder, out, a, b, batch_shape, a_batch_strides, b_batch_strides);
 }
 
 void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -143,6 +154,28 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto [a_transposed, lda, a] = check_transpose(encoder, s, a_pre);
   auto [b_transposed, ldb, b] = check_transpose(encoder, s, b_pre);
 
+  /////////////////////////////////////////////////////////////////////////////
+  // Dispatch to GEMM with epilogue or AddMM
+
+  if (beta_ == 1 && c.strides(-1) == 1 && c.data_size() == out.shape(-1)) {
+    out.set_data(allocator::malloc(out.nbytes()));
+    gemm_and_bias(
+        encoder,
+        M,
+        N,
+        K,
+        a_transposed,
+        lda,
+        b_transposed,
+        ldb,
+        out,
+        a,
+        b,
+        c,
+        alpha_);
+    return;
+  }
+
   int64_t ldc;
   {
     auto stx = c.strides()[c.ndim() - 2];
@@ -184,9 +217,9 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
 
   /////////////////////////////////////////////////////////////////////////////
-  // Invoke cublasLt
+  // Invoke cublasLt with AddMM settings
 
-  cu::Matmul matmul(
+  CublasGemm gemm(
       cu::device(s.device),
       a.dtype(),
       a_transposed,
@@ -202,12 +235,7 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
       a_batch_strides.back(),
       b_batch_strides.back(),
       c_batch_strides.back());
-
+  gemm.run(
-  if ((batch_count / batch_shape.back()) == 1) {
-    matmul.run(encoder, out, a, b, c, alpha_, beta_);
-    return;
-  }
-  matmul.run_batched(
       encoder,
       out,
       a,

mlx/backend/cuda/no_cuda.cpp

@@ -1,11 +1,47 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/cuda/cuda.h"
+#include "mlx/fast.h"
 
-namespace mlx::core::cu {
+namespace mlx::core {
+
+namespace cu {
 
 bool is_available() {
   return false;
 }
 
-} // namespace mlx::core::cu
+} // namespace cu
+
+namespace fast {
+
+CustomKernelFunction cuda_kernel(
+    const std::string&,
+    const std::vector<std::string>&,
+    const std::vector<std::string>&,
+    const std::string&,
+    const std::string&,
+    bool,
+    int) {
+  throw std::runtime_error("[cuda_kernel] No CUDA back-end.");
+}
+
+std::vector<array> precompiled_cuda_kernel(
+    const std::string&,
+    const std::string&,
+    const std::vector<array>&,
+    const std::vector<Shape>&,
+    const std::vector<Dtype>&,
+    const std::vector<ScalarArg>&,
+    std::tuple<int, int, int>,
+    std::tuple<int, int, int>,
+    int shared_memory,
+    std::optional<float> init_value,
+    bool ensure_row_contiguous,
+    StreamOrDevice) {
+  throw std::runtime_error("[cuda_kernel] No CUDA back-end.");
+}
+
+} // namespace fast
+
+} // namespace mlx::core

mlx/backend/cuda/primitives.cpp

@@ -24,8 +24,6 @@ namespace mlx::core {
   }
 
 NO_GPU(BlockMaskedMM)
-NO_GPU(DynamicSlice)
-NO_GPU(DynamicSliceUpdate)
 NO_GPU(FFT)
 NO_GPU(GatherMM)
 NO_GPU(GatherQMM)
@@ -41,12 +39,7 @@ NO_GPU(Cholesky)
 NO_GPU_MULTI(Eig)
 NO_GPU_MULTI(Eigh)
 
-namespace fast {
-NO_GPU_MULTI(CustomKernel)
-} // namespace fast
-
 namespace distributed {
-NO_GPU_MULTI(AllReduce)
 NO_GPU_MULTI(AllGather)
 NO_GPU_MULTI(Send)
 NO_GPU_MULTI(Recv)

mlx/backend/cuda/quantized/quantized.cpp

@@ -46,10 +46,10 @@ inline array ensure_row_contiguous_matrix(
 
 } // namespace
 
-void fast::AffineQuantize::eval_gpu(
+void fast::Quantize::eval_gpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
-  nvtx3::scoped_range r("AffineQuantize::eval_gpu");
+  nvtx3::scoped_range r("Quantize::eval_gpu");
   auto& s = stream();
   auto& d = cu::device(s.device);
   auto& enc = d.get_command_encoder(s);

mlx/backend/cuda/rope.cu

@@ -103,15 +103,21 @@ template <typename T, bool traditional, bool forward, int N = 4>
 __device__ void rope_impl(
     const T* in,
     T* out,
-    int offset,
+    const int* offset,
     float inv_freq,
     float scale,
     const cuda::std::array<int64_t, 3> strides,
     const cuda::std::array<int64_t, 3> out_strides,
-    int64_t n_batch,
+    int64_t offset_stride,
+    int n_head,
     uint3 pos,
     uint3 dims) {
-  float L = scale * static_cast<float>(pos.y + offset);
+  auto n_head_up = N * ((n_head + N - 1) / N);
+  auto head_idx = static_cast<int>((pos.z * N) % n_head_up);
+  auto batch_idx = (pos.z * N) / n_head_up;
+  auto batch_offset = offset[batch_idx * offset_stride];
+  float L = scale * static_cast<float>(pos.y + batch_offset);
+  auto mat_idx = batch_idx * n_head + head_idx;
 
   // Compute costheta, sintheta
   float theta = L * inv_freq;
@@ -123,20 +129,19 @@ __device__ void rope_impl(
   size_t out_index_1, out_index_2;
   if (traditional) {
     out_index_1 = 2 * pos.x * out_strides[2] + pos.y * out_strides[1] +
-        N * pos.z * out_strides[0];
+        mat_idx * out_strides[0];
     out_index_2 = out_index_1 + 1;
     in_index_1 =
-        2 * pos.x * strides[2] + pos.y * strides[1] + N * pos.z * strides[0];
+        2 * pos.x * strides[2] + pos.y * strides[1] + mat_idx * strides[0];
     in_index_2 = in_index_1 + strides[2];
   } else {
     out_index_1 = pos.x * out_strides[2] + pos.y * out_strides[1] +
-        N * pos.z * out_strides[0];
+        mat_idx * out_strides[0];
     out_index_2 = out_index_1 + dims.x * out_strides[2];
-    in_index_1 =
+    in_index_1 = pos.x * strides[2] + pos.y * strides[1] + mat_idx * strides[0];
-        pos.x * strides[2] + pos.y * strides[1] + N * pos.z * strides[0];
     in_index_2 = in_index_1 + dims.x * strides[2];
   }
-  for (int i = 0; i < N && pos.z * N + i < n_batch; ++i) {
+  for (int i = 0; i < N && head_idx + i < n_head; ++i) {
     // Read and write the output
     float x1 = static_cast<float>(in[in_index_1]);
     float x2 = static_cast<float>(in[in_index_2]);
@@ -167,7 +172,8 @@ __global__ void rope(
     float base,
     const __grid_constant__ cuda::std::array<int64_t, 3> strides,
     const __grid_constant__ cuda::std::array<int64_t, 3> out_strides,
-    int64_t n_batch,
+    int64_t offset_stride,
+    int n_head,
     uint3 dims) {
   uint3 pos = make_uint3(
       blockIdx.x * blockDim.x + threadIdx.x,
@@ -182,12 +188,13 @@ __global__ void rope(
   rope_impl<T, traditional, forward>(
       in,
       out,
-      *offset,
+      offset,
       inv_freq,
       scale,
       strides,
       out_strides,
-      n_batch,
+      offset_stride,
+      n_head,
       pos,
       dims);
 }
@@ -202,7 +209,8 @@ __global__ void rope_freqs(
     float base,
     const __grid_constant__ cuda::std::array<int64_t, 3> strides,
     const __grid_constant__ cuda::std::array<int64_t, 3> out_strides,
-    int64_t n_batch,
+    int64_t offset_stride,
+    int n_head,
     uint3 dims,
     int64_t freq_stride) {
   uint3 pos = make_uint3(
@@ -217,12 +225,13 @@ __global__ void rope_freqs(
   rope_impl<T, traditional, forward>(
       in,
       out,
-      *offset,
+      offset,
       inv_freq,
       scale,
       strides,
       out_strides,
-      n_batch,
+      offset_stride,
+      n_head,
       pos,
       dims);
 }
@@ -245,23 +254,28 @@ void RoPE::eval_gpu(
   auto& offset = inputs[1];
   auto& out = outputs[0];
 
-  if (in.ndim() < 3) {
-    throw std::runtime_error("[RoPE] Input must have at least 3 dimensions");
-  }
-
   cuda::std::array<int64_t, 3> strides;
   cuda::std::array<int64_t, 3> out_strides;
   bool donated = false;
   int ndim = in.ndim();
-  int dispatch_ndim = in.ndim();
+
+  int B = in.shape(0);
+  int T = in.shape(-2);
+  int D = in.shape(-1);
+  size_t mat_size = T * D;
+  int dispatch_ndim = ndim;
   while (in.shape(-dispatch_ndim) == 1 && dispatch_ndim > 3) {
     dispatch_ndim--;
   }
-  size_t mat_size = in.shape(-2) * in.shape(-1);
+
+  int N = 1;
+  for (int i = 1; i < (ndim - 2); ++i) {
+    N *= in.shape(i);
+  }
 
   // We apply rope to less that the whole vector so copy to output and then
   // apply in-place.
-  if (dims_ < in.shape(-1)) {
+  if (dims_ < D) {
     donated = true;
     auto ctype =
         (in.flags().row_contiguous) ? CopyType::Vector : CopyType::General;
@@ -302,7 +316,7 @@ void RoPE::eval_gpu(
   out_strides[2] = out.strides()[ndim - 1];
 
   // Some flags to help us dispatch below
-  bool single = in.flags().row_contiguous && (mat_size == in.shape(-1));
+  bool single = in.flags().row_contiguous && B == 1 && T == 1;
   bool with_freqs = inputs.size() == 3;
 
   auto& encoder = cu::get_command_encoder(s);
@@ -319,7 +333,7 @@ void RoPE::eval_gpu(
         if (single && !with_freqs) {
           auto kernel =
               cu::rope_single<DataType, traditional.value, forward.value>;
-          uint2 dims = make_uint2(dims_ / 2, in.size() / mat_size);
+          uint2 dims = make_uint2(dims_ / 2, N);
           auto [grid, block] = get_grid_and_block(dims.x, dims.y, 1);
           encoder.add_kernel_node(
               kernel,
@@ -336,7 +350,7 @@ void RoPE::eval_gpu(
         } else if (single) {
           auto kernel =
               cu::rope_single_freqs<DataType, traditional.value, forward.value>;
-          uint2 dims = make_uint2(dims_ / 2, in.size() / mat_size);
+          uint2 dims = make_uint2(dims_ / 2, N);
           auto [grid, block] = get_grid_and_block(dims.x, dims.y, 1);
           encoder.add_kernel_node(
               kernel,
@@ -354,10 +368,14 @@ void RoPE::eval_gpu(
         } else if (with_freqs) {
           auto kernel =
               cu::rope_freqs<DataType, traditional.value, forward.value>;
-          uint3 dims =
+          int n_per_thread = 4;
-              make_uint3(dims_ / 2, in.shape(-2), in.size() / mat_size);
+          uint32_t dimz = B * ((N + n_per_thread - 1) / n_per_thread);
-          dims.z = (dims.z + 3) / 4;
+          uint3 dims = make_uint3(dims_ / 2, T, dimz);
           auto [grid, block] = get_grid_and_block(dims.x, dims.y, dims.z);
+          int64_t offset_stride = 0;
+          if (inputs[1].ndim() > 0) {
+            offset_stride = inputs[1].strides()[0];
+          }
           encoder.add_kernel_node(
               kernel,
               grid,
@@ -371,15 +389,20 @@ void RoPE::eval_gpu(
               std::log2(base_),
               strides,
               out_strides,
-              in.size() / mat_size,
+              offset_stride,
+              N,
               dims,
               inputs[2].strides(0));
         } else {
           auto kernel = cu::rope<DataType, traditional.value, forward.value>;
-          uint3 dims =
+          int n_per_thread = 4;
-              make_uint3(dims_ / 2, in.shape(-2), in.size() / mat_size);
+          uint32_t dimz = B * ((N + n_per_thread - 1) / n_per_thread);
-          dims.z = (dims.z + 3) / 4;
+          uint3 dims = make_uint3(dims_ / 2, T, dimz);
           auto [grid, block] = get_grid_and_block(dims.x, dims.y, dims.z);
+          int64_t offset_stride = 0;
+          if (inputs[1].ndim() > 0) {
+            offset_stride = inputs[1].strides()[0];
+          }
           encoder.add_kernel_node(
               kernel,
               grid,
@@ -392,7 +415,8 @@ void RoPE::eval_gpu(
               std::log2(base_),
               strides,
               out_strides,
-              in.size() / mat_size,
+              offset_stride,
+              N,
               dims);
         }
       });

mlx/backend/cuda/scaled_dot_product_attention.cu

@@ -4,23 +4,16 @@
 #include "mlx/backend/cuda/device/config.h"
 #include "mlx/backend/cuda/device/utils.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/backend/cuda/lru_cache.h"
 #include "mlx/backend/gpu/copy.h"
 #include "mlx/dtype_utils.h"
 #include "mlx/fast_primitives.h"
+#include "mlx/transforms_impl.h"
 
-// cudnn_frontend.h redefines this macro.
-#undef CHECK_CUDA_ERROR
-
-#include <cudnn_frontend.h>
-#include <fmt/format.h>
 #include <nvtx3/nvtx3.hpp>
 
 #include <cooperative_groups.h>
 #include <cooperative_groups/reduce.h>
 
-namespace fe = cudnn_frontend;
-
 namespace mlx::core {
 
 namespace cu {
@@ -52,6 +45,7 @@ __global__ void kernel_sdpav_1pass(
     const T* K,
     const T* V,
     T* O,
+    const T* sinks,
     __grid_constant__ const AttnParams params) {
   constexpr int BN = 32;
   constexpr int BD = 32;
@@ -71,7 +65,7 @@ __global__ void kernel_sdpav_1pass(
   __shared__ U max_scores[BN];
   __shared__ U sum_exp_scores[BN];
 
-  const U scale_log2 = params.scale * 1.44269504089f;
+  const U scale_log2 = params.scale * M_LOG2E;
 
   auto block = cg::this_thread_block();
   auto warp = cg::tiled_partition<32>(block);
@@ -114,8 +108,12 @@ __global__ void kernel_sdpav_1pass(
     o[i] = 0.f;
   }
 
-  U max_score = -INFINITY;
+  U max_score = Limits<U>::finite_min();
   U sum_exp_score = 0.f;
+  if (sinks && warp_idx == 0) {
+    max_score = M_LOG2E * static_cast<U>(sinks[head_idx]);
+    sum_exp_score = 1.f;
+  }
 
   // For each key
   for (int i = kv_seq_idx; i < params.kL; i += BN) {
@@ -173,7 +171,7 @@ __global__ void kernel_sdpav_1pass(
   U factor = exp2f(max_score - new_max);
   sum_exp_score =
       cg::reduce(warp, sum_exp_scores[lane_idx] * factor, cg::plus<U>());
-  sum_exp_score = __frcp_rn(sum_exp_score);
+  sum_exp_score = sum_exp_score == 0 ? 0 : __frcp_rn(sum_exp_score);
 
   // Now we need to aggregate all the outputs
   PRAGMA_LOOP_UNROLL
@@ -199,6 +197,7 @@ __global__ void kernel_sdpav_2pass_1(
     const T* Q,
     const T* K,
     const T* V,
+    const T* sinks,
     float* partials,
     float* sums,
     float* maxs,
@@ -274,8 +273,12 @@ __global__ void kernel_sdpav_2pass_1(
     o[i] = 0.f;
   }
 
-  U max_score = -1e9;
+  U max_score = Limits<U>::finite_min();
   U sum_exp_score = 0.f;
+  if (sinks && warp_idx == 0 && block_idx == 0) {
+    max_score = M_LOG2E * static_cast<U>(sinks[head_idx]);
+    sum_exp_score = 1.f;
+  }
 
   // For each key
   for (int i = kv_seq_idx; i < params.kL; i += blocks * BN) {
@@ -416,7 +419,7 @@ __global__ void kernel_sdpav_2pass_2(
   U new_max = cg::reduce(warp, max_score, cg::greater<U>());
   U factor = exp2f(max_score - new_max);
   U sum_exp_score = cg::reduce(warp, sums[lane_idx] * factor, cg::plus<U>());
-  sum_exp_score = __frcp_rn(sum_exp_score);
+  sum_exp_score = sum_exp_score == 0 ? 0 : __frcp_rn(sum_exp_score);
 
   PRAGMA_LOOP_UNROLL
   for (int i = 0; i < v_per_thread; i++) {
@@ -469,10 +472,14 @@ void sdpa_vector_1pass_fallback(
     const array& v,
     const float scale,
     array& o,
-    bool do_causal_ = false) {
+    bool do_causal,
+    const std::optional<array>& sinks) {
   encoder.set_input_array(q);
   encoder.set_input_array(k);
   encoder.set_input_array(v);
+  if (sinks) {
+    encoder.set_input_array(*sinks);
+  }
   encoder.set_output_array(o);
 
   cu::AttnParams params{
@@ -495,7 +502,7 @@ void sdpa_vector_1pass_fallback(
   dim3 block_dim(1024, 1, 1);
 
   dispatch_float_types(o.dtype(), "kernel_sdpav_1pass", [&](auto type_tag) {
-    dispatch_bool(do_causal_, [&](auto do_causal) {
+    dispatch_bool(do_causal, [&](auto do_causal) {
       dispatch_headdim(params.D, [&](auto headdim) {
         using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
 
@@ -510,6 +517,7 @@ void sdpa_vector_1pass_fallback(
             k.data<DataType>(),
             v.data<DataType>(),
             o.data<DataType>(),
+            sinks ? (*sinks).data<DataType>() : nullptr,
             params);
       });
     });
@@ -524,7 +532,8 @@ void sdpa_vector_2pass_fallback(
     const array& v,
     const float scale,
     array& o,
-    bool do_causal_ = false) {
+    bool do_causal,
+    const std::optional<array>& sinks) {
   cu::AttnParams params{
       /* int B = */ q.shape(0),
       /* int H = */ q.shape(1),
@@ -565,7 +574,7 @@ void sdpa_vector_2pass_fallback(
   encoder.add_temporary(maxs);
 
   dispatch_float_types(o.dtype(), "kernel_sdpav_2pass", [&](auto type_tag) {
-    dispatch_bool(do_causal_, [&](auto do_causal) {
+    dispatch_bool(do_causal, [&](auto do_causal) {
       dispatch_headdim(params.D, [&](auto headdim) {
         using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
 
@@ -576,6 +585,10 @@ void sdpa_vector_2pass_fallback(
           encoder.set_input_array(q);
           encoder.set_input_array(k);
           encoder.set_input_array(v);
+          if (sinks) {
+            encoder.set_input_array(*sinks);
+          }
+
           encoder.set_output_array(intermediate);
           encoder.set_output_array(sums);
           encoder.set_output_array(maxs);
@@ -591,6 +604,7 @@ void sdpa_vector_2pass_fallback(
               q.data<DataType>(),
               k.data<DataType>(),
               v.data<DataType>(),
+              sinks ? (*sinks).data<DataType>() : nullptr,
               intermediate.data<float>(),
               sums.data<float>(),
               maxs.data<float>(),
@@ -633,306 +647,19 @@ void sdpa_vector_fallback(
     const array& v,
     const float scale,
     array& o,
-    bool do_causal_ = false) {
+    bool do_causal,
+    const std::optional<array>& sinks) {
   int kL = k.shape(2);
 
   if (kL > 1024) {
     return sdpa_vector_2pass_fallback(
-        s, encoder, q, k, v, scale, o, do_causal_);
+        s, encoder, q, k, v, scale, o, do_causal, sinks);
   } else {
     return sdpa_vector_1pass_fallback(
-        s, encoder, q, k, v, scale, o, do_causal_);
+        s, encoder, q, k, v, scale, o, do_causal, sinks);
   }
 }
 
-struct SDPACacheKey {
-  int device_id;
-  fe::DataType_t cudnn_type;
-
-  int B;
-  int H;
-  int D;
-
-  int qL;
-  int kL;
-
-  int gqa_factor;
-  float scale;
-
-  int64_t Q_strides[3];
-  int64_t K_strides[3];
-  int64_t V_strides[3];
-  int64_t O_strides[3];
-
-  bool generate_stats;
-  bool causal_mask;
-};
-
-auto& sdpa_cache() {
-  static LRUBytesKeyCache<SDPACacheKey, std::shared_ptr<fe::graph::Graph>>
-      cache(
-          /* capacity */ 128);
-  return cache;
-}
-
-#define Q_UID 1
-#define K_UID 2
-#define V_UID 3
-#define O_UID 4
-#define STATS_UID 5
-
-std::shared_ptr<fe::graph::Graph> get_sdpa_forward_graph(
-    cu::CommandEncoder& encoder,
-    const SDPACacheKey& cache_key) {
-  // Check if graph has already been fully built
-  if (auto it = sdpa_cache().find(cache_key); it != sdpa_cache().end()) {
-    return it->second;
-  }
-
-  // Set up new graph
-  auto graph = std::make_shared<fe::graph::Graph>();
-
-  graph->set_io_data_type(cache_key.cudnn_type)
-      .set_intermediate_data_type(fe::DataType_t::FLOAT)
-      .set_compute_data_type(fe::DataType_t::FLOAT);
-
-  auto Q = graph->tensor(
-      fe::graph::Tensor_attributes()
-          .set_name("Q")
-          .set_uid(Q_UID)
-          .set_dim({cache_key.B, cache_key.H, cache_key.qL, cache_key.D})
-          .set_stride(
-              {cache_key.Q_strides[0],
-               cache_key.Q_strides[1],
-               cache_key.Q_strides[2],
-               1}));
-
-  int h_kv = cache_key.H / cache_key.gqa_factor;
-  auto K =
-      graph->tensor(fe::graph::Tensor_attributes()
-                        .set_name("K")
-                        .set_uid(K_UID)
-                        .set_dim({cache_key.B, h_kv, cache_key.kL, cache_key.D})
-                        .set_stride(
-                            {cache_key.K_strides[0],
-                             cache_key.K_strides[1],
-                             cache_key.V_strides[2],
-                             1}));
-
-  auto V =
-      graph->tensor(fe::graph::Tensor_attributes()
-                        .set_name("V")
-                        .set_uid(V_UID)
-                        .set_dim({cache_key.B, h_kv, cache_key.kL, cache_key.D})
-                        .set_stride(
-                            {cache_key.V_strides[0],
-                             cache_key.V_strides[1],
-                             cache_key.V_strides[2],
-                             1}));
-
-  auto sdpa_options = fe::graph::SDPA_attributes()
-                          .set_name("flash_attention")
-                          .set_is_inference(!cache_key.generate_stats)
-                          .set_attn_scale(cache_key.scale);
-
-  if (cache_key.causal_mask && cache_key.qL > 1) {
-    sdpa_options.set_diagonal_alignment(fe::DiagonalAlignment_t::TOP_LEFT)
-        .set_diagonal_band_right_bound(0);
-  }
-
-  auto [O, Stats] = graph->sdpa(Q, K, V, sdpa_options);
-
-  O->set_output(true)
-      .set_uid(O_UID)
-      .set_dim({cache_key.B, cache_key.H, cache_key.qL, cache_key.D})
-      .set_stride(
-          {cache_key.O_strides[0],
-           cache_key.O_strides[1],
-           cache_key.O_strides[2],
-           1});
-
-  if (cache_key.generate_stats) {
-    Stats->set_output(true)
-        .set_data_type(fe::DataType_t::FLOAT)
-        .set_uid(STATS_UID);
-  }
-
-  // Build and Validate cudnn graph
-
-  auto handle = encoder.device().cudnn_handle();
-
-  // cuDNN only supports native CUDA graphs for sdpa in 9.6 or above.
-  if (cudnnGetVersion() < 90600) {
-    auto build_status = graph->build(handle, {fe::HeurMode_t::A});
-    if (!build_status.is_good()) {
-      throw std::runtime_error(
-          "Unable to build cudnn graph for attention."
-          " Failed with message: " +
-          build_status.get_message());
-    }
-
-  } else {
-    auto val_status = graph->validate();
-    auto op_status = graph->build_operation_graph(handle);
-
-    auto plan_stauts =
-        graph->create_execution_plans({cudnn_frontend::HeurMode_t::A});
-    if (!plan_stauts.is_good()) {
-      throw std::runtime_error(
-          "Unable to create exec plan for cudnn attention."
-          " Failed with message: " +
-          plan_stauts.get_message());
-    }
-
-    graph->select_behavior_notes(
-        {cudnn_frontend::BehaviorNote_t::SUPPORTS_CUDA_GRAPH_NATIVE_API});
-
-    auto support_status = graph->check_support(handle);
-    if (!support_status.is_good()) {
-      throw std::runtime_error(
-          "No cuda graph support for cudnn attention."
-          " Failed with message: " +
-          support_status.get_message());
-    }
-
-    auto build_status = graph->build_plans(handle);
-    if (!build_status.is_good()) {
-      throw std::runtime_error(
-          "Unable to build cudnn graph for attention."
-          " Failed with message: " +
-          build_status.get_message());
-    }
-  }
-
-  auto [it, _] = sdpa_cache().emplace(cache_key, graph);
-
-  return it->second;
-}
-
-inline fe::DataType_t dtype_to_cudnn_type(Dtype dtype) {
-  switch (dtype) {
-    case int8:
-      return fe::DataType_t::INT8;
-    case int32:
-      return fe::DataType_t::INT32;
-    case uint8:
-      return fe::DataType_t::UINT8;
-    case float16:
-      return fe::DataType_t::HALF;
-    case bfloat16:
-      return fe::DataType_t::BFLOAT16;
-    case float32:
-      return fe::DataType_t::FLOAT;
-    case float64:
-      return fe::DataType_t::DOUBLE;
-    default:
-      throw std::runtime_error(fmt::format(
-          "Unsupported dtype in SDPA: {}.", dtype_to_string(dtype)));
-  }
-}
-
-void sdpa_cudnn(
-    const Stream& s,
-    cu::CommandEncoder& encoder,
-    const array& q,
-    const array& k,
-    const array& v,
-    const float scale,
-    array& o,
-    bool do_causal_ = false) {
-  encoder.set_input_array(q);
-  encoder.set_input_array(k);
-  encoder.set_input_array(v);
-  encoder.set_output_array(o);
-
-  auto cudnn_type = dtype_to_cudnn_type(q.dtype());
-
-  int B = q.shape(0);
-  int H = q.shape(1);
-  int D = q.shape(3);
-  int gqa_factor = q.shape(1) / k.shape(1);
-
-  int qL = q.shape(2);
-  int kL = k.shape(2);
-
-  SDPACacheKey cache_key{
-      /* int device_id = */ encoder.device().cuda_device(),
-      /* fe::DataType_t cudnn_type = */ cudnn_type,
-
-      /* int B = */ B,
-      /* int H = */ H,
-      /* int D = */ D,
-
-      /* int qL = */ qL,
-      /* int kL = */ kL,
-
-      /* int gqa_factor = */ gqa_factor,
-      /* float scale = */ scale,
-
-      /* int64_t Q_strides[3] = */ {q.strides(0), q.strides(1), q.strides(2)},
-      /* int64_t K_strides[3] = */ {k.strides(0), k.strides(1), k.strides(2)},
-      /* int64_t V_strides[3] = */ {v.strides(0), v.strides(1), v.strides(2)},
-      /* int64_t O_strides[3] = */ {o.strides(0), o.strides(1), o.strides(2)},
-
-      /* bool generate_stats = */ false,
-      /* bool causal_mask = */ do_causal_};
-
-  auto graph = get_sdpa_forward_graph(encoder, cache_key);
-
-  int64_t workspace_size = 0;
-  auto workspace_status = graph->get_workspace_size(workspace_size);
-  if (!workspace_status.is_good()) {
-    throw std::runtime_error("Unable to get workspace for cudnn attention.");
-  }
-
-  array workspace(
-      allocator::malloc(workspace_size), {int(workspace_size)}, uint8);
-  auto workspace_ptr = workspace.data<void>();
-
-  std::unordered_map<int64_t, void*> variant_pack = {
-      {Q_UID, const_cast<void*>(q.data<void>())},
-      {K_UID, const_cast<void*>(k.data<void>())},
-      {V_UID, const_cast<void*>(v.data<void>())},
-      {O_UID, o.data<void>()}};
-
-  auto handle = encoder.device().cudnn_handle();
-  cudnnSetStream(handle, encoder.stream());
-
-  // cuDNN only supports native CUDA graphs for sdpa in 9.6 or above.
-  if (cudnnGetVersion() < 90600) {
-    auto capture = encoder.capture_context();
-    auto exec_status = graph->execute(handle, variant_pack, workspace_ptr);
-
-    if (!exec_status.is_good()) {
-      capture.discard = true;
-      throw std::runtime_error(
-          "Unable to execute cudnn attention."
-          " Failed with message: " +
-          exec_status.get_message());
-    }
-  } else {
-    cudaGraph_t cu_graph;
-    cudaGraphCreate(&cu_graph, 0);
-
-    std::unique_ptr<cudaGraph_t, void (*)(cudaGraph_t*)> graph_freer(
-        &cu_graph, [](cudaGraph_t* p) { cudaGraphDestroy(*p); });
-
-    auto cu_graph_status = graph->populate_cuda_graph(
-        handle, variant_pack, workspace_ptr, cu_graph);
-
-    if (!cu_graph_status.is_good()) {
-      throw std::runtime_error(
-          "Unable to add cuda graph for cudnn attention."
-          " Failed with message: " +
-          cu_graph_status.get_message());
-    }
-
-    encoder.add_graph_node(cu_graph);
-  }
-
-  encoder.add_temporary(workspace);
-}
-
 } // namespace
 
 namespace fast {
@@ -945,6 +672,9 @@ bool ScaledDotProductAttention::use_fallback(
     bool has_arr_mask,
     bool do_causal,
     Stream s) {
+  if (detail::in_grad_tracing()) {
+    return true;
+  }
   if (s.device == Device::cpu) {
     return true;
   }
@@ -960,15 +690,7 @@ bool ScaledDotProductAttention::use_fallback(
   const bool supported_vector_config =
       sdpa_supported_head_dim && query_sequence_length < 4;
 
-  auto& cu_device = cu::device(s.device);
+  const bool supported_config = supported_vector_config;
-
-  const bool supported_matrix_config = query_sequence_length > 4 &&
-      cu_device.compute_capability_major() >= 8 &&
-      query_sequence_length == key_sequence_length &&
-      (q.dtype() == float16 || q.dtype() == bfloat16);
-
-  const bool supported_config =
-      (supported_matrix_config || supported_vector_config);
 
   return has_arr_mask || !supported_config;
 }
@@ -990,7 +712,7 @@ void ScaledDotProductAttention::eval_gpu(
 
   // Define some copy functions to ensure the layout of the inputs is as
   // expected.
-  copies.reserve(3);
+  copies.reserve(inputs.size());
   auto copy_unless = [&copies, &s](
                          auto predicate, const array& arr) -> const array& {
     if (!predicate(arr)) {
@@ -1002,10 +724,16 @@ void ScaledDotProductAttention::eval_gpu(
     }
   };
 
+  // Checks that the headdim dimension has stride 1.
   auto is_matrix_contiguous = [](const array& arr) {
     return arr.strides(-1) == 1;
   };
 
+  std::optional<array> sinks = std::nullopt;
+  if (has_sinks_) {
+    sinks = copy_unless(is_matrix_contiguous, inputs.back());
+  }
+
   // We are in vector mode ie single query
   if (q_pre.shape(2) < 4) {
     auto q_copy_unless = [](const array& arr) {
@@ -1043,10 +771,6 @@ void ScaledDotProductAttention::eval_gpu(
     const auto& k = copy_unless(kv_copy_unless, k_pre);
     const auto& v = copy_unless(kv_copy_unless, v_pre);
 
-    for (const auto& cp : copies) {
-      encoder.add_temporary(cp);
-    }
-
     // Donate the query if possible
     if (q.is_donatable() && q.flags().row_contiguous && q.size() == o.size()) {
       o.copy_shared_buffer(q);
@@ -1055,53 +779,31 @@ void ScaledDotProductAttention::eval_gpu(
       int64_t str_oH = o.shape(3);
       int64_t str_oL = o.shape(1) * str_oH;
       int64_t str_oB = o.shape(2) * str_oL;
-      size_t data_size = o.shape(0) * str_oB;
 
       array::Flags flags{
           /* bool contiguous = */ 1,
-          /* bool row_contiguous = */ 0,
+          /* bool row_contiguous = */ o.shape(2) == 1,
-          /* bool col_contiguous = */ 0,
+          /* bool col_contiguous = */ o.size() == o.shape(3),
       };
 
       o.set_data(
           allocator::malloc(o.nbytes()),
-          data_size,
+          o.size(),
           {str_oB, str_oH, str_oL, str_oD},
           flags);
     }
 
-    return sdpa_vector_fallback(s, encoder, q, k, v, scale_, o, do_causal_);
-  }
-
-  // Full attention mode
-  else {
-    const auto& q = copy_unless(is_matrix_contiguous, q_pre);
-    const auto& k = copy_unless(is_matrix_contiguous, k_pre);
-    const auto& v = copy_unless(is_matrix_contiguous, v_pre);
-
     for (const auto& cp : copies) {
       encoder.add_temporary(cp);
     }
 
-    int64_t str_oD = 1;
+    return sdpa_vector_fallback(
-    int64_t str_oH = o.shape(3);
+        s, encoder, q, k, v, scale_, o, do_causal_, sinks);
-    int64_t str_oL = o.shape(1) * str_oH;
+  }
-    int64_t str_oB = o.shape(2) * str_oL;
-    size_t data_size = o.shape(0) * str_oB;
 
-    array::Flags flags{
+  // Full attention mode should never reach here
-        /* bool contiguous = */ 1,
+  else {
-        /* bool row_contiguous = */ 0,
+    throw std::runtime_error("Doesn't support matrix yet.");
-        /* bool col_contiguous = */ 0,
-    };
-
-    o.set_data(
-        allocator::malloc(o.nbytes()),
-        data_size,
-        {str_oB, str_oH, str_oL, str_oD},
-        flags);
-
-    return sdpa_cudnn(s, encoder, q, k, v, scale_, o, do_causal_);
   }
 }
 

mlx/backend/cuda/slicing.cpp

@@ -1,8 +1,11 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/common/slicing.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/jit_module.h"
 #include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/gpu/slicing.h"
+#include "mlx/dtype_utils.h"
 
 #include <numeric>
 
@@ -27,8 +30,7 @@ void concatenate_gpu(
   flags.row_contiguous = false;
   flags.col_contiguous = false;
   flags.contiguous = false;
-  // TODO: Handle concurrent outputs:
+  auto concurrent = cu::get_command_encoder(s).concurrent_context();
-  // https://github.com/ml-explore/mlx/pull/2145#discussion_r2070753816
   for (int i = 0; i < inputs.size(); i++) {
     array out_slice(inputs[i].shape(), out.dtype(), nullptr, {});
     size_t data_offset = strides[axis] * sizes[i];
@@ -38,4 +40,71 @@ void concatenate_gpu(
   }
 }
 
+array compute_dynamic_offset(
+    const array& indices,
+    const Strides& strides,
+    const std::vector<int>& axes,
+    const Stream& s) {
+  Dtype dtype = indices.dtype();
+  int nidx = axes.size();
+
+  std::string module_name =
+      fmt::format("compute_dynamic_offset_{}_{}", dtype_to_string(dtype), nidx);
+  std::string kernel_name = fmt::format(
+      "mlx::core::cu::compute_dynamic_offset<{}, {}>",
+      dtype_to_cuda_type(dtype),
+      nidx);
+
+  cu::JitModule& mod = cu::get_jit_module(s.device, module_name, [&]() {
+    std::string source = R"(
+        #include "mlx/backend/cuda/device/utils.cuh"
+
+        namespace mlx::core::cu {
+
+        template <typename T, int NIDX>
+        __global__ void compute_dynamic_offset(
+            const T* indices,
+            int64_t* offset,
+            const __grid_constant__ Strides strides,
+            const __grid_constant__ cuda::std::array<int, NIDX> axes) {
+          int64_t acc = 0;
+          #pragma unroll
+          for (int i = 0; i < NIDX; ++i) {
+            acc += indices[i] * strides[axes[i]];
+          }
+          *offset = acc;
+        }
+
+        } // namespace mlx::core::cu
+    )";
+    return std::make_tuple(false, std::move(source), std::vector{kernel_name});
+  });
+
+  // Prepare output.
+  array offset({1}, int64, nullptr, {});
+  bool donate = indices.is_donatable() &&
+      (indices.data_size() * indices.itemsize()) >= offset.itemsize();
+  if (donate) {
+    offset.copy_shared_buffer(indices);
+  } else {
+    offset.set_data(allocator::malloc(offset.itemsize()));
+  }
+
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.add_temporary(offset);
+  encoder.set_input_array(indices);
+  encoder.set_output_array(offset);
+
+  cu::KernelArgs args;
+  args.append(indices);
+  args.append(offset);
+  args.append_ndim(strides);
+  args.append(axes);
+
+  auto kernel = mod.get_kernel(kernel_name);
+  encoder.add_kernel_node(kernel, 1, 1, 0, args.args());
+
+  return offset;
+}
+
 } // namespace mlx::core

mlx/backend/cuda/sort.cu

@@ -1,6 +1,5 @@
 // Copyright © 2025 Apple Inc.
 
-#include "mlx/backend/common/utils.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/backend/gpu/copy.h"
@@ -10,7 +9,7 @@
 #include <nvtx3/nvtx3.hpp>
 #include <thrust/device_ptr.h>
 #include <thrust/transform.h>
-#include <cub/device/device_segmented_sort.cuh>
+#include <cub/device/device_segmented_radix_sort.cuh>
 
 #include <cassert>
 
@@ -80,7 +79,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
         encoder.add_temporary(discard);
 
         size_t size;
-        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortPairs(
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedRadixSort::SortPairs(
             nullptr,
             size,
             in.data<Type>(),
@@ -91,6 +90,8 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             in.data_size() / nsort,
             offsets,
             offsets + 1,
+            0,
+            sizeof(Type) * 8,
             stream));
 
         array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
@@ -105,7 +106,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             thrust::device_pointer_cast(indices.data<uint32_t>()),
             ModOp<uint32_t>{static_cast<uint32_t>(nsort)});
 
-        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortPairs(
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedRadixSort::SortPairs(
             temp.data<void>(),
             size,
             in.data<Type>(),
@@ -116,10 +117,12 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             in.data_size() / nsort,
             offsets,
             offsets + 1,
+            0,
+            sizeof(Type) * 8,
             stream));
       } else {
         size_t size;
-        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortKeys(
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedRadixSort::SortKeys(
             nullptr,
             size,
             in.data<Type>(),
@@ -128,6 +131,8 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             in.data_size() / nsort,
             offsets,
             offsets + 1,
+            0,
+            sizeof(Type) * 8,
             stream));
 
         array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
@@ -135,7 +140,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
 
         // Start capturing after allocations
         auto capture = encoder.capture_context();
-        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortKeys(
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedRadixSort::SortKeys(
             temp.data<void>(),
             size,
             in.data<Type>(),
@@ -144,6 +149,8 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             in.data_size() / nsort,
             offsets,
             offsets + 1,
+            0,
+            sizeof(Type) * 8,
             stream));
       }
     } else {

mlx/backend/cuda/ternary.cu

@@ -39,52 +39,98 @@ ternary_v(const bool* a, const T* b, const T* c, T* out, IdxT size) {
   }
 }
 
-template <typename Op, typename T, typename IdxT, int NDIM>
+template <typename Op, typename T, typename IdxT, int NDIM, int N_READS>
 __global__ void ternary_g_nd(
     const bool* a,
     const T* b,
     const T* c,
     T* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> c_strides) {
-  IdxT index = cg::this_grid().thread_rank();
+  auto block = cg::this_thread_block();
-  if (index < size) {
+  auto grid = cg::this_grid();
+  IdxT index_rest =
+      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
+  if (index_rest >= size_rest) {
+    return;
+  }
+
+  auto shape_x = shape[NDIM - 1];
+  auto a_stride_x = a_strides[NDIM - 1];
+  auto b_stride_x = b_strides[NDIM - 1];
+  auto c_stride_x = c_strides[NDIM - 1];
+  IdxT index_x =
+      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
   auto [a_idx, b_idx, c_idx] = elem_to_loc_nd<NDIM>(
-        index,
+      index_rest * shape_x,
       shape.data(),
       a_strides.data(),
       b_strides.data(),
       c_strides.data());
-    out[index] = Op{}(a[a_idx], b[b_idx], c[c_idx]);
+  auto a_vec =
+      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, false);
+  auto b_vec =
+      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, T(0));
+  auto c_vec =
+      load_vector<N_READS>(c + c_idx, index_x, shape_x, c_stride_x, T(0));
+
+  AlignedVector<T, N_READS> out_vec;
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    out_vec[i] = Op{}(a_vec[i], b_vec[i], c_vec[i]);
   }
+  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
-template <typename Op, typename T, typename IdxT>
+template <typename Op, typename T, typename IdxT, int N_READS>
 __global__ void ternary_g(
     const bool* a,
     const T* b,
     const T* c,
     T* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides a_strides,
     const __grid_constant__ Strides b_strides,
     const __grid_constant__ Strides c_strides,
     int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
+  auto block = cg::this_thread_block();
-  if (index < size) {
+  auto grid = cg::this_grid();
+  IdxT index_rest =
+      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
+  if (index_rest >= size_rest) {
+    return;
+  }
+
+  auto shape_x = shape[ndim - 1];
+  auto a_stride_x = a_strides[ndim - 1];
+  auto b_stride_x = b_strides[ndim - 1];
+  auto c_stride_x = c_strides[ndim - 1];
+  IdxT index_x =
+      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
   auto [a_idx, b_idx, c_idx] = elem_to_loc(
-        index,
+      index_rest * shape_x,
       shape.data(),
       a_strides.data(),
       b_strides.data(),
       c_strides.data(),
       ndim);
-    out[index] = Op{}(a[a_idx], b[b_idx], c[c_idx]);
+  auto a_vec =
+      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, false);
+  auto b_vec =
+      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, T(0));
+  auto c_vec =
+      load_vector<N_READS>(c + c_idx, index_x, shape_x, c_stride_x, T(0));
+
+  AlignedVector<T, N_READS> out_vec;
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    out_vec[i] = Op{}(a_vec[i], b_vec[i], c_vec[i]);
   }
+  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
 } // namespace cu
@@ -123,36 +169,55 @@ void ternary_op_gpu_inplace(
             auto& b_strides = strides[1];
             auto& c_strides = strides[2];
             int ndim = shape.size();
+            int work_per_thread = 1;
+            auto dim0 = ndim > 0 ? shape.back() : 1;
+            auto rest = out.size() / dim0;
+            if (dim0 >= 4) {
+              work_per_thread = 4;
+            }
+            dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
+            auto block_dims = get_block_dims(dim0, rest, 1);
+            uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
+            uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
+
             if (ndim <= 3) {
               dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                auto [num_blocks, block_dims] = get_launch_args(out, large());
+                auto kernel =
+                    cu::ternary_g_nd<Op, DType, IdxT, dims_constant(), 1>;
+                if (work_per_thread == 4) {
+                  kernel =
+                      cu::ternary_g_nd<Op, DType, IdxT, dims_constant(), 4>;
+                }
                 encoder.add_kernel_node(
-                    cu::ternary_g_nd<Op, DType, IdxT, dims_constant()>,
+                    kernel,
-                    num_blocks,
+                    {num_blocks_x, num_blocks_y},
                     block_dims,
                     0,
                     a.data<bool>(),
                     b.data<DType>(),
                     c.data<DType>(),
                     out.data<DType>(),
-                    out.size(),
+                    rest,
                     const_param<dims_constant()>(shape),
                     const_param<dims_constant()>(a_strides),
                     const_param<dims_constant()>(b_strides),
                     const_param<dims_constant()>(c_strides));
               });
             } else {
-              auto [num_blocks, block_dims] = get_launch_args(out, large());
+              auto kernel = cu::ternary_g<Op, DType, IdxT, 1>;
+              if (work_per_thread == 4) {
+                kernel = cu::ternary_g<Op, DType, IdxT, 4>;
+              }
               encoder.add_kernel_node(
-                  cu::ternary_g<Op, DType, IdxT>,
+                  kernel,
-                  num_blocks,
+                  {num_blocks_x, num_blocks_y},
                   block_dims,
                   0,
                   a.data<bool>(),
                   b.data<DType>(),
                   c.data<DType>(),
                   out.data<DType>(),
-                  out.data_size(),
+                  rest,
                   const_param(shape),
                   const_param(a_strides),
                   const_param(b_strides),

mlx/backend/cuda/unary.cu

@@ -37,19 +37,36 @@ __global__ void unary_v(const In* in, Out* out, IdxT size) {
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void unary_g(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides strides,
     int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
+  auto block = cg::this_thread_block();
-  if (index < size) {
+  auto grid = cg::this_grid();
-    auto idx = elem_to_loc(index, shape.data(), strides.data(), ndim);
+  IdxT index_rest =
-    out[index] = Op{}(in[idx]);
+      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
+  if (index_rest >= size_rest) {
+    return;
   }
+
+  auto shape_x = shape[ndim - 1];
+  auto stride_x = strides[ndim - 1];
+  IdxT index_x =
+      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  auto idx =
+      elem_to_loc(index_rest * shape_x, shape.data(), strides.data(), ndim);
+  auto in_vec =
+      load_vector<N_READS>(in + idx, index_x, shape_x, stride_x, In(0));
+  AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    out_vec[i] = Op{}(in_vec[i]);
+  }
+  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
 template <typename Op, typename In, typename Out>
@@ -127,8 +144,7 @@ void unary_op_gpu_inplace(
           using OutType = cuda_type_t<CTYPE_OUT>;
           if (contig) {
             using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-            // TODO: Choose optimized value based on type size.
+            constexpr int N_READS = 16 / sizeof(OutType);
-            constexpr int N_READS = 4;
             auto [num_blocks, block_dims] = get_launch_args(
                 out.data_size(), out.shape(), out.strides(), large, N_READS);
             encoder.add_kernel_node(
@@ -142,18 +158,30 @@ void unary_op_gpu_inplace(
           } else {
             using IdxT = std::conditional_t<large(), int64_t, int32_t>;
             auto [shape, strides] = collapse_contiguous_dims(in);
-            auto [num_blocks, block_dims] = get_launch_args(out, large);
+            auto ndim = shape.size();
+            int work_per_thread = 1;
+            auto kernel = cu::unary_g<Op, InType, OutType, IdxT, 1>;
+            auto dim0 = ndim > 0 ? shape.back() : 1;
+            auto rest = out.size() / dim0;
+            if (dim0 >= 4) {
+              kernel = cu::unary_g<Op, InType, OutType, IdxT, 4>;
+              work_per_thread = 4;
+            }
+            dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
+            auto block_dims = get_block_dims(dim0, rest, 1);
+            uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
+            uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
             encoder.add_kernel_node(
-                cu::unary_g<Op, InType, OutType, IdxT>,
+                kernel,
-                num_blocks,
+                {num_blocks_x, num_blocks_y},
                 block_dims,
                 0,
                 in.data<InType>(),
                 out.data<OutType>(),
-                out.data_size(),
+                rest,
                 const_param(shape),
                 const_param(strides),
-                shape.size());
+                ndim);
           }
         });
       } else {

mlx/backend/cuda/unary/CMakeLists.txt

@@ -0,0 +1,34 @@
+target_sources(
+  mlx
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/abs.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arccos.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arccosh.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arcsin.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arcsinh.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arctan.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arctanh.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/bitwise_invert.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/ceil.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/conjugate.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/cos.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/cosh.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/erf.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/erf_inv.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/exp.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/expm1.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/floor.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/imag.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/log.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/log1p.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/logical_not.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/negative.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/real.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/round.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/sigmoid.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/sign.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/sin.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/sinh.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/sqrt.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/square.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/tan.cu
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/tanh.cu)

mlx/backend/cuda/unary/abs.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/unary/unary.cuh"
+
+namespace mlx::core {
+UNARY_GPU(Abs)
+} // namespace mlx::core

mlx/backend/cuda/unary/arccos.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/unary/unary.cuh"
+
+namespace mlx::core {
+UNARY_GPU(ArcCos)
+} // namespace mlx::core

mlx/backend/cuda/unary/arccosh.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/unary/unary.cuh"
+
+namespace mlx::core {
+UNARY_GPU(ArcCosh)
+} // namespace mlx::core

mlx/backend/cuda/unary/arcsin.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/unary/unary.cuh"
+
+namespace mlx::core {
+UNARY_GPU(ArcSin)
+} // namespace mlx::core

mlx/backend/cuda/unary/arcsinh.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/unary/unary.cuh"
+
+namespace mlx::core {
+UNARY_GPU(ArcSinh)
+} // namespace mlx::core

mlx/backend/cuda/unary/arctan.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/unary/unary.cuh"
+
+namespace mlx::core {
+UNARY_GPU(ArcTan)
+} // namespace mlx::core

mlx/backend/cuda/unary/arctanh.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/unary/unary.cuh"
+
+namespace mlx::core {
+UNARY_GPU(ArcTanh)
+} // namespace mlx::core

mlx/backend/cuda/unary/bitwise_invert.cu

@@ -0,0 +1,7 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/unary/unary.cuh"
+
+namespace mlx::core {
+UNARY_GPU(BitwiseInvert)
+} // namespace mlx::core